diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_tensor([367, 367])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_tensor([367, 367])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ce8e3d73753d4ba71b3938b070c9eeb8a098ac9c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_tensor([367, 367])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot depicts a simple physics setup in a craft room. Under the warm, focused light of an unseen desk lamp, a smooth, cylindrical wooden dowel begins to roll steadily down a simple ramp made of pine planks. When it reaches the bottom, it collides with a second, identical dowel resting on the worn surface of a wooden workbench. The impact decisively sends the second dowel sliding to the right, while the first deflects and continues its slow roll to the left. The camera remains fixed as both dowels lose momentum and gently come to rest, now separated and still at the bottom of the ramp. The realistic lighting casts soft, long shadows, accentuating the natural grain of the wood and creating an atmosphere of quiet, focused, hands-on experimentation.", + "seed": 17358, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_367_seed_17358_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": -0.0351984778419137, + "vlm_logprob": -0.0351984778419137, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.0351984778419137, + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050195686519145966 + ], + "opposite_logprob_tokens": [ + -3.5501956939697266 + ], + "logprob_reward": -0.050195686519145966, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.050195686519145966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20732033252716064, + 0.1469905972480774, + 0.13137011229991913, + 0.061210013926029205, + 0.04479262977838516, + 0.11512096971273422, + 0.0774088129401207, + 0.0867108553647995, + 0.06825987994670868, + 0.06081583723425865 + ], + "credit_map": [ + 0.025196054950356483, + 0.02744252048432827, + 0.012741131708025932, + 0.02291887067258358, + 0.027852552011609077, + 0.025002796202898026, + 0.03571445122361183, + 0.030451955273747444, + 0.009366164915263653, + 0.03102266602218151, + 0.006351357325911522, + 0.00862470269203186, + 0.013978107832372189, + 0.011032265610992908, + 0.03156355395913124, + 0.03505177050828934, + 0.012720334343612194, + 0.01685461960732937, + 0.008298089727759361, + 0.012300964444875717, + 0.02490762062370777, + 0.02268732339143753, + 0.020802507176995277, + 0.01279864739626646, + 0.006915197242051363, + 0.010738917626440525, + 0.003983089234679937, + 0.004654146730899811, + 0.011418825015425682, + 0.007956570014357567, + 0.00972523633390665, + 0.005818038713186979, + 0.0045488448813557625, + 0.007332668174058199, + 0.003739306703209877, + 0.004788438323885202, + 0.009268058463931084, + 0.0038018759805709124, + 0.006104490254074335, + 0.005208950489759445, + 0.010434852913022041, + 0.01650962233543396, + 0.006939609535038471, + 0.010756261646747589, + 0.01376305054873228, + 0.014382239431142807, + 0.026502955704927444, + 0.015832383185625076, + 0.008137612603604794, + 0.011170519515872002, + 0.005473928991705179, + 0.0076926411129534245, + 0.010825596749782562, + 0.008324225433170795, + 0.014919507317245007, + 0.010864783078432083, + 0.01272436324506998, + 0.015563126653432846, + 0.00913618877530098, + 0.008213486522436142, + 0.01132869254797697, + 0.00903585460036993, + 0.008977096527814865, + 0.011732044629752636, + 0.007984722964465618, + 0.015390099957585335, + 0.0034986429382115602, + 0.00441204896196723, + 0.009966610930860043, + 0.005897908937186003, + 0.011951751075685024, + 0.009158089756965637, + 0.005650288425385952, + 0.007670493796467781, + 0.0030474888626486063, + 0.006891419645398855, + 0.009370082058012486, + 0.010330627672374249, + 0.008168225176632404, + 0.009687207639217377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.050195686519145966 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01424113567918539 + ], + "opposite_logprob_tokens": [ + -4.2642412185668945 + ], + "logprob_reward": -0.01424113567918539, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01424113567918539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11121440678834915, + 0.1864878237247467, + 0.0562036894261837, + 0.02864678204059601, + 0.06562153249979019, + 0.14681287109851837, + 0.07351775467395782, + 0.13334104418754578, + 0.15592801570892334, + 0.04222604259848595 + ], + "credit_map": [ + 0.016024595126509666, + 0.01118582021445036, + 0.007739349268376827, + 0.009669413790106773, + 0.012389264069497585, + 0.01682046614587307, + 0.02567332051694393, + 0.01171218790113926, + 0.009279074147343636, + 0.0462271124124527, + 0.007291646674275398, + 0.0117601677775383, + 0.016594067215919495, + 0.01740393415093422, + 0.044100575149059296, + 0.03383124619722366, + 0.006148260552436113, + 0.008207197301089764, + 0.003728351788595319, + 0.004872910678386688, + 0.012896579690277576, + 0.006675963755697012, + 0.007954557426273823, + 0.005719868000596762, + 0.0032057566568255424, + 0.003555201692506671, + 0.002137020230293274, + 0.003822009777650237, + 0.003965935669839382, + 0.0032376027666032314, + 0.004767120815813541, + 0.003956136293709278, + 0.006328236777335405, + 0.00979906227439642, + 0.0029918141663074493, + 0.005413421895354986, + 0.01242414116859436, + 0.009807958267629147, + 0.01223338395357132, + 0.0066235242411494255, + 0.008390103466808796, + 0.01732947863638401, + 0.00648311385884881, + 0.012273880653083324, + 0.023516613990068436, + 0.015457961708307266, + 0.04361167550086975, + 0.01975005492568016, + 0.006350482814013958, + 0.01148403063416481, + 0.004309538751840591, + 0.005546113476157188, + 0.010091128759086132, + 0.0091580580919981, + 0.014520506374537945, + 0.012057902291417122, + 0.010919558815658092, + 0.020395947620272636, + 0.009018242359161377, + 0.009543171152472496, + 0.020827285945415497, + 0.014830442145466805, + 0.02116633951663971, + 0.026640059426426888, + 0.008481602184474468, + 0.0233460683375597, + 0.005117154214531183, + 0.009955862537026405, + 0.028425151482224464, + 0.013894520699977875, + 0.0468650721013546, + 0.01984260417521, + 0.003856740426272154, + 0.005351732950657606, + 0.0025049527175724506, + 0.003823003266006708, + 0.006323798559606075, + 0.005645866971462965, + 0.008303497917950153, + 0.006416447926312685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.01424113567918539 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003571093082428 + ], + "opposite_logprob_tokens": [ + -2.695035696029663 + ], + "logprob_reward": -0.07003571093082428, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003571093082428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07420473545789719, + 0.46091175079345703, + 0.03417110815644264, + 0.04322543740272522, + 0.03159555420279503, + 0.06059955060482025, + 0.04047894850373268, + 0.1440926343202591, + 0.07557593286037445, + 0.03514433279633522 + ], + "credit_map": [ + 0.011111694388091564, + 0.008735285140573978, + 0.005383045878261328, + 0.006224996875971556, + 0.008750692009925842, + 0.012077766470611095, + 0.013670782558619976, + 0.008250472135841846, + 0.023618174716830254, + 0.13766229152679443, + 0.013093474321067333, + 0.01615992933511734, + 0.04326615482568741, + 0.023870471864938736, + 0.10612168163061142, + 0.09711962938308716, + 0.0034714958164840937, + 0.004157527349889278, + 0.0021920003928244114, + 0.002869933843612671, + 0.006846013944596052, + 0.003966478165239096, + 0.006018614396452904, + 0.004649044014513493, + 0.0037821552250534296, + 0.0069830529391765594, + 0.0029035110492259264, + 0.003380636917427182, + 0.007844919338822365, + 0.006057496648281813, + 0.0071051958948373795, + 0.005168473348021507, + 0.002654501236975193, + 0.004106201697140932, + 0.001669483259320259, + 0.004546326585114002, + 0.003994489088654518, + 0.004011205397546291, + 0.0064256684854626656, + 0.004187684506177902, + 0.003850267268717289, + 0.006386739667505026, + 0.003048536367714405, + 0.004517575725913048, + 0.009723096154630184, + 0.006411650683730841, + 0.01823350600898266, + 0.00842818059027195, + 0.004153532441705465, + 0.004857392515987158, + 0.0025304662995040417, + 0.0025141029618680477, + 0.005849351640790701, + 0.004449415951967239, + 0.010356776416301727, + 0.005767914932221174, + 0.01474404614418745, + 0.02307145670056343, + 0.00523601146414876, + 0.00808858685195446, + 0.024088367819786072, + 0.014697390608489513, + 0.031187284737825394, + 0.022979501634836197, + 0.0035173639189451933, + 0.014352665282785892, + 0.0018948349170386791, + 0.004911412484943867, + 0.016225852072238922, + 0.004952583461999893, + 0.020437296479940414, + 0.009283928200602531, + 0.0029874967876821756, + 0.004083920735865831, + 0.002667095046490431, + 0.003289415268227458, + 0.006606520153582096, + 0.0032343976199626923, + 0.007767822593450546, + 0.004507668782025576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.07003571093082428 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023334892466664314 + ], + "opposite_logprob_tokens": [ + -3.7733349800109863 + ], + "logprob_reward": -0.023334892466664314, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023334892466664314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16874849796295166, + 0.1528162956237793, + 0.05114705488085747, + 0.027598731219768524, + 0.04340788722038269, + 0.053680215030908585, + 0.03795940801501274, + 0.08178362250328064, + 0.3246214985847473, + 0.05823684111237526 + ], + "credit_map": [ + 0.018564492464065552, + 0.01775810681283474, + 0.011006693355739117, + 0.019390851259231567, + 0.019953355193138123, + 0.02476319670677185, + 0.04109242558479309, + 0.016219360753893852, + 0.013726935721933842, + 0.021767927333712578, + 0.007659854367375374, + 0.016363708302378654, + 0.017563965171575546, + 0.017478125169873238, + 0.031864747405052185, + 0.026391008868813515, + 0.006744292564690113, + 0.005918785929679871, + 0.003802961204200983, + 0.004803386516869068, + 0.010692796669900417, + 0.006135950796306133, + 0.007766463328152895, + 0.005282411817461252, + 0.0029915766790509224, + 0.00433700904250145, + 0.0022389590740203857, + 0.003023262368515134, + 0.0044059790670871735, + 0.0024873060174286366, + 0.0050478787161409855, + 0.0030667567625641823, + 0.003546071005985141, + 0.0049262456595897675, + 0.003055848414078355, + 0.004667569883167744, + 0.007701484020799398, + 0.00528172729536891, + 0.009075595997273922, + 0.005153343081474304, + 0.004939752630889416, + 0.007288704626262188, + 0.0029214657843112946, + 0.005489321891218424, + 0.008184881880879402, + 0.007098871748894453, + 0.011079944670200348, + 0.0066772629506886005, + 0.004276371095329523, + 0.00440506637096405, + 0.0024778787046670914, + 0.0036088675260543823, + 0.005540273617953062, + 0.004753117449581623, + 0.008700028993189335, + 0.004197798203676939, + 0.006248330697417259, + 0.013330725952982903, + 0.005452911369502544, + 0.005356121808290482, + 0.014347957447171211, + 0.008499106392264366, + 0.014284062199294567, + 0.014264398254454136, + 0.026415474712848663, + 0.04448120296001434, + 0.015877312049269676, + 0.024542544037103653, + 0.05012648180127144, + 0.03645165637135506, + 0.06757301092147827, + 0.05915376916527748, + 0.0038531136233359575, + 0.00940951518714428, + 0.0022373704705387354, + 0.004963803570717573, + 0.01147453673183918, + 0.007207565940916538, + 0.009217946790158749, + 0.009872984141111374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.023334892466664314 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818496361374855 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.01818496361374855, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818496361374855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3760600984096527, + 0.14654190838336945, + 0.04024830460548401, + 0.05120133236050606, + 0.03958021476864815, + 0.09731046855449677, + 0.054051704704761505, + 0.1153959259390831, + 0.05399860814213753, + 0.025611398741602898 + ], + "credit_map": [ + 0.05341000109910965, + 0.04039496183395386, + 0.01811847649514675, + 0.033775970339775085, + 0.04726853221654892, + 0.0564383901655674, + 0.08878755569458008, + 0.037866245955228806, + 0.008786790072917938, + 0.022134874016046524, + 0.005900912918150425, + 0.008222350850701332, + 0.019911151379346848, + 0.01892715133726597, + 0.026143306866288185, + 0.03651539608836174, + 0.003980021458119154, + 0.0059664249420166016, + 0.00229851552285254, + 0.0032075156923383474, + 0.00775529071688652, + 0.005118026398122311, + 0.006551092024892569, + 0.005371426697820425, + 0.005133696366101503, + 0.012624663300812244, + 0.003004815662279725, + 0.00434931181371212, + 0.009744418784976006, + 0.003971418831497431, + 0.006393441464751959, + 0.005979571957141161, + 0.0031640466768294573, + 0.005827962886542082, + 0.0022920507472008467, + 0.004002527333796024, + 0.006632031407207251, + 0.005667142570018768, + 0.007988293655216694, + 0.004006166476756334, + 0.005619099363684654, + 0.01081529539078474, + 0.003911986015737057, + 0.007260733284056187, + 0.016145018860697746, + 0.0069447350688278675, + 0.03329628333449364, + 0.013317330740392208, + 0.004979479592293501, + 0.0065996237099170685, + 0.0031646357383579016, + 0.005139064509421587, + 0.008033698424696922, + 0.005762152373790741, + 0.012148357927799225, + 0.008224695920944214, + 0.010674446821212769, + 0.024806128814816475, + 0.004015551880002022, + 0.006398747209459543, + 0.020071877166628838, + 0.009058882482349873, + 0.0221530981361866, + 0.018217213451862335, + 0.0037117241881787777, + 0.009011195041239262, + 0.0014086096780374646, + 0.0029286150820553303, + 0.010109523311257362, + 0.0050689708441495895, + 0.011341869831085205, + 0.010418102145195007, + 0.002020228886976838, + 0.0035500752273947, + 0.0015660267090424895, + 0.002254317281767726, + 0.004293491132557392, + 0.0035205932799726725, + 0.004797434434294701, + 0.0036092333029955626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01818496361374855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot depicts a simple physics setup in a craft room. Under the warm, focused light of an unseen desk lamp, a smooth, cylindrical wooden dowel begins to roll steadily down a simple ramp made of pine planks. When it reaches the bottom, it collides with a second, identical dowel resting on the worn surface of a wooden workbench. The impact decisively sends the second dowel sliding to the right, while the first deflects and continues its slow roll to the left. The camera remains fixed as both dowels lose momentum and gently come to rest, now separated and still at the bottom of the ramp. The realistic lighting casts soft, long shadows, accentuating the natural grain of the wood and creating an atmosphere of quiet, focused, hands-on experimentation.", + "seed": 17359, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_367_seed_17359_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": -0.03514457792043686, + "vlm_logprob": -0.03514457792043686, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03514457792043686, + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07275896519422531 + ], + "opposite_logprob_tokens": [ + -2.947758913040161 + ], + "logprob_reward": -0.07275896519422531, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07275896519422531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1438223272562027, + 0.3019064962863922, + 0.03773696348071098, + 0.07508698850870132, + 0.10651028156280518, + 0.05528527870774269, + 0.07695453613996506, + 0.07612386345863342, + 0.07573871314525604, + 0.05083455145359039 + ], + "credit_map": [ + 0.011089188046753407, + 0.021304424852132797, + 0.012003069743514061, + 0.02051771804690361, + 0.009532155469059944, + 0.012833154760301113, + 0.03516591712832451, + 0.021376704797148705, + 0.017527727410197258, + 0.05538026988506317, + 0.018896283581852913, + 0.017953936010599136, + 0.059684138745069504, + 0.0534493587911129, + 0.04590383917093277, + 0.03311093896627426, + 0.00352416536770761, + 0.005091544706374407, + 0.005407185759395361, + 0.005313345231115818, + 0.0029405567329376936, + 0.002995478454977274, + 0.007159240078181028, + 0.0053054471500217915, + 0.0051654428243637085, + 0.007887739688158035, + 0.014868589118123055, + 0.012029853649437428, + 0.006554412189871073, + 0.005887627135962248, + 0.010336511768400669, + 0.012356812134385109, + 0.010667731985449791, + 0.013354616239666939, + 0.011533653363585472, + 0.015602762810885906, + 0.012617822736501694, + 0.011656398884952068, + 0.01680164225399494, + 0.01427565235644579, + 0.0032641554716974497, + 0.007916099391877651, + 0.008628797717392445, + 0.008569441735744476, + 0.006777375005185604, + 0.004819528665393591, + 0.009475680999457836, + 0.005834197159856558, + 0.005911856424063444, + 0.010421277955174446, + 0.010243616066873074, + 0.009631035849452019, + 0.007522362284362316, + 0.008380399085581303, + 0.013977647759020329, + 0.010866343975067139, + 0.005242145154625177, + 0.008490918204188347, + 0.009493802674114704, + 0.00817556120455265, + 0.010752862319350243, + 0.011411016806960106, + 0.013074195012450218, + 0.009483360685408115, + 0.005401039496064186, + 0.00902929250150919, + 0.008129121735692024, + 0.008521666750311852, + 0.011083945631980896, + 0.01005506981164217, + 0.013682493939995766, + 0.009836078621447086, + 0.003564647864550352, + 0.006303819827735424, + 0.008417150005698204, + 0.009873565286397934, + 0.0044587887823581696, + 0.006084639113396406, + 0.006553429178893566, + 0.0055785151198506355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07275896519422531 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02985270321369171 + ], + "opposite_logprob_tokens": [ + -3.5298526287078857 + ], + "logprob_reward": -0.02985270321369171, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02985270321369171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14715054631233215, + 0.1206841766834259, + 0.11015360057353973, + 0.13450554013252258, + 0.05385889485478401, + 0.019916828721761703, + 0.1732691377401352, + 0.07525060325860977, + 0.06884044408798218, + 0.09637025743722916 + ], + "credit_map": [ + 0.012440292164683342, + 0.027923032641410828, + 0.017557131126523018, + 0.02262650616466999, + 0.010279862210154533, + 0.01211576908826828, + 0.022922778502106667, + 0.021285170689225197, + 0.008526737801730633, + 0.024359771981835365, + 0.008211198262870312, + 0.012812647968530655, + 0.01592867076396942, + 0.012643902562558651, + 0.027133304625749588, + 0.011067948304116726, + 0.009448678232729435, + 0.015569650568068027, + 0.01181967742741108, + 0.012030789628624916, + 0.008158313110470772, + 0.0075235371477901936, + 0.02961243875324726, + 0.015990514308214188, + 0.010549850761890411, + 0.017078373581171036, + 0.020326079800724983, + 0.020224591717123985, + 0.013237508945167065, + 0.010898959822952747, + 0.02025437168776989, + 0.02193579636514187, + 0.005639104172587395, + 0.008836107328534126, + 0.0063491747714579105, + 0.006681673228740692, + 0.004520166665315628, + 0.006032044067978859, + 0.007828143425285816, + 0.007972477935254574, + 0.0015530847012996674, + 0.0032776810694485903, + 0.004303103778511286, + 0.003258730284869671, + 0.0018363796407356858, + 0.0018395335646346211, + 0.002791154198348522, + 0.0010571627644822001, + 0.013007685542106628, + 0.026523470878601074, + 0.023343827575445175, + 0.027555588632822037, + 0.015056698583066463, + 0.016476722434163094, + 0.034499458968639374, + 0.016805678606033325, + 0.005253125447779894, + 0.010193916969001293, + 0.010846502147614956, + 0.00897818710654974, + 0.010212266817688942, + 0.009092219173908234, + 0.011328469030559063, + 0.009345916099846363, + 0.005084732081741095, + 0.009217245504260063, + 0.015087207779288292, + 0.00890808179974556, + 0.005044037010520697, + 0.007179821375757456, + 0.00981414970010519, + 0.008505163714289665, + 0.0075895292684435844, + 0.014117635786533356, + 0.012410300783813, + 0.011539192870259285, + 0.011892859824001789, + 0.010172209702432156, + 0.012859283946454525, + 0.015789248049259186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.02985270321369171 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308167099952698 + ], + "opposite_logprob_tokens": [ + -3.168081760406494 + ], + "logprob_reward": -0.04308167099952698, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308167099952698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13681671023368835, + 0.08867798000574112, + 0.06270682066679001, + 0.06737972795963287, + 0.052850134670734406, + 0.021209605038166046, + 0.2276388704776764, + 0.11563675850629807, + 0.10933417826890945, + 0.1177491694688797 + ], + "credit_map": [ + 0.008903102949261665, + 0.019750244915485382, + 0.01839400827884674, + 0.017896253615617752, + 0.011481765657663345, + 0.012292132712900639, + 0.027384554967284203, + 0.020714640617370605, + 0.00557664642110467, + 0.01640857383608818, + 0.0078655406832695, + 0.008571894839406013, + 0.012459808960556984, + 0.01090225949883461, + 0.01717705838382244, + 0.009716199710965157, + 0.00564391165971756, + 0.008736935444176197, + 0.008037558756768703, + 0.0076680430211126804, + 0.004791643004864454, + 0.0037704382557421923, + 0.015420284122228622, + 0.008638003841042519, + 0.005846274551004171, + 0.009469417855143547, + 0.01062324084341526, + 0.011260660365223885, + 0.0055144596844911575, + 0.0045487890020012856, + 0.009557489305734634, + 0.010559397749602795, + 0.005963385105133057, + 0.006261279806494713, + 0.00562609639018774, + 0.006638514809310436, + 0.004742657300084829, + 0.006744662299752235, + 0.00775811867788434, + 0.009115420281887054, + 0.0017345055239275098, + 0.0037221903912723064, + 0.004018712323158979, + 0.002867424627766013, + 0.0024624948855489492, + 0.0018029414350166917, + 0.003124469658359885, + 0.0014768668916076422, + 0.01719464734196663, + 0.03491148725152016, + 0.0306719858199358, + 0.03511954843997955, + 0.02026684582233429, + 0.022325605154037476, + 0.04560030996799469, + 0.0215484369546175, + 0.009588992223143578, + 0.016886597499251366, + 0.015079529955983162, + 0.013441921211779118, + 0.01194316241890192, + 0.01349785178899765, + 0.015961186960339546, + 0.019237520173192024, + 0.008784153498709202, + 0.011734811589121819, + 0.019200347363948822, + 0.013427847996354103, + 0.0081516457721591, + 0.01313485112041235, + 0.015943536534905434, + 0.018956979736685753, + 0.009555617347359657, + 0.01570611260831356, + 0.021755401045084, + 0.014143667183816433, + 0.015623650513589382, + 0.013156795874238014, + 0.01612786203622818, + 0.011680057272315025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04308167099952698 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15963806211948395, + 0.115276038646698, + 0.07416056841611862, + 0.11672979593276978, + 0.22236652672290802, + 0.04046504572033882, + 0.03813999891281128, + 0.09503510594367981, + 0.10720077902078629, + 0.03098807856440544 + ], + "credit_map": [ + 0.010364037938416004, + 0.019637247547507286, + 0.012102531269192696, + 0.021308809518814087, + 0.013240679167211056, + 0.01702631637454033, + 0.03646116331219673, + 0.029497282579541206, + 0.007985211908817291, + 0.024125458672642708, + 0.00817771628499031, + 0.007717679720371962, + 0.017858149483799934, + 0.01555540505796671, + 0.02049686200916767, + 0.013359551317989826, + 0.00624651275575161, + 0.009742704220116138, + 0.008209100924432278, + 0.00918186828494072, + 0.006075640209019184, + 0.007695665117353201, + 0.01632370799779892, + 0.010685368441045284, + 0.00952840968966484, + 0.01398154441267252, + 0.019218003377318382, + 0.021091999486088753, + 0.009564018808305264, + 0.012375099584460258, + 0.014111923053860664, + 0.016858795657753944, + 0.020944980904459953, + 0.030206482857465744, + 0.020615261048078537, + 0.0328049436211586, + 0.03132141754031181, + 0.034387536346912384, + 0.033456478267908096, + 0.01862943358719349, + 0.0032970195170491934, + 0.004856819286942482, + 0.005814559757709503, + 0.006401245482265949, + 0.00497044064104557, + 0.005314299836754799, + 0.006614262238144875, + 0.003196400823071599, + 0.0033641825430095196, + 0.004836539272218943, + 0.004233948420733213, + 0.006307351402938366, + 0.003225449239835143, + 0.003020753152668476, + 0.007889809086918831, + 0.0052619678899645805, + 0.007781158201396465, + 0.0109106982126832, + 0.010587157681584358, + 0.011147919110953808, + 0.008689949288964272, + 0.011649486608803272, + 0.014398803934454918, + 0.019869931042194366, + 0.008975626900792122, + 0.012650569900870323, + 0.011301495134830475, + 0.012908220291137695, + 0.011129772290587425, + 0.012348742224276066, + 0.01730220764875412, + 0.020584149286150932, + 0.0026642130687832832, + 0.004296471830457449, + 0.004409391433000565, + 0.003986992873251438, + 0.0029394926968961954, + 0.0033638563472777605, + 0.004670327994972467, + 0.004657335113734007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.006746135652065277 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "opposite_logprob_tokens": [ + -3.7732834815979004 + ], + "logprob_reward": -0.02328341454267502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328341454267502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17716084420681, + 0.08645161986351013, + 0.1026720479130745, + 0.05489782243967056, + 0.10574720799922943, + 0.043289780616760254, + 0.09850574284791946, + 0.16822536289691925, + 0.1262587606906891, + 0.03679085895419121 + ], + "credit_map": [ + 0.011118579655885696, + 0.020756322890520096, + 0.012036529369652271, + 0.01952819712460041, + 0.018338242545723915, + 0.018453998491168022, + 0.03814568743109703, + 0.03878327086567879, + 0.00560083519667387, + 0.012963252142071724, + 0.00688276207074523, + 0.008503826335072517, + 0.009680897928774357, + 0.006810091435909271, + 0.02348216436803341, + 0.012527787126600742, + 0.012399345636367798, + 0.013103973120450974, + 0.009785809554159641, + 0.008501531556248665, + 0.011108361184597015, + 0.014906722120940685, + 0.017740434035658836, + 0.015125862322747707, + 0.005181232932955027, + 0.0072283390909433365, + 0.006418669130653143, + 0.008272538892924786, + 0.006343325134366751, + 0.006307146046310663, + 0.007569042965769768, + 0.0075775254517793655, + 0.011286242865025997, + 0.013986453413963318, + 0.011879730969667435, + 0.014803043566644192, + 0.011285144835710526, + 0.013388218358159065, + 0.016027312725782394, + 0.013091056607663631, + 0.003331104526296258, + 0.004985737148672342, + 0.005474942736327648, + 0.006372417788952589, + 0.00508916936814785, + 0.0050210291519761086, + 0.008371587842702866, + 0.004643786698579788, + 0.00813300535082817, + 0.01727779395878315, + 0.010830421932041645, + 0.011896960437297821, + 0.008838492445647717, + 0.01009514182806015, + 0.017353486269712448, + 0.014080436900258064, + 0.01432760339230299, + 0.01781580038368702, + 0.016855204477906227, + 0.019683485850691795, + 0.014373826794326305, + 0.021735547110438347, + 0.027194533497095108, + 0.036239344626665115, + 0.01092851348221302, + 0.01431390643119812, + 0.010682451538741589, + 0.01511460542678833, + 0.010708412155508995, + 0.014634865336120129, + 0.021698512136936188, + 0.028177481144666672, + 0.003016406437382102, + 0.0046049063093960285, + 0.003632265143096447, + 0.004751029424369335, + 0.0031039398163557053, + 0.0039047186728566885, + 0.006157344672828913, + 0.007620244286954403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02328341454267502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot depicts a simple physics setup in a craft room. Under the warm, focused light of an unseen desk lamp, a smooth, cylindrical wooden dowel begins to roll steadily down a simple ramp made of pine planks. When it reaches the bottom, it collides with a second, identical dowel resting on the worn surface of a wooden workbench. The impact decisively sends the second dowel sliding to the right, while the first deflects and continues its slow roll to the left. The camera remains fixed as both dowels lose momentum and gently come to rest, now separated and still at the bottom of the ramp. The realistic lighting casts soft, long shadows, accentuating the natural grain of the wood and creating an atmosphere of quiet, focused, hands-on experimentation.", + "seed": 17360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_367_seed_17360_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": -0.030081637762486936, + "vlm_logprob": -0.030081637762486936, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.030081637762486936, + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038765400648117065 + ], + "opposite_logprob_tokens": [ + -3.6637654304504395 + ], + "logprob_reward": -0.038765400648117065, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.038765400648117065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1621713787317276, + 0.1334148496389389, + 0.023125817999243736, + 0.018794313073158264, + 0.25603532791137695, + 0.12997476756572723, + 0.09068597108125687, + 0.031118085607886314, + 0.06791350245475769, + 0.08676597476005554 + ], + "credit_map": [ + 0.01873389072716236, + 0.01557747833430767, + 0.031205788254737854, + 0.022184964269399643, + 0.015354689210653305, + 0.01878487318754196, + 0.019129332154989243, + 0.02120036818087101, + 0.010252016596496105, + 0.01338550727814436, + 0.03493175655603409, + 0.011812577024102211, + 0.008766056038439274, + 0.009089335799217224, + 0.014919301494956017, + 0.0302583035081625, + 0.002890242962166667, + 0.0030515592079609632, + 0.003870793618261814, + 0.002448566257953644, + 0.002345220884308219, + 0.002569659613072872, + 0.002737955888733268, + 0.00321182026527822, + 0.002936597680673003, + 0.0033567596692591906, + 0.0022316647227853537, + 0.001771013718098402, + 0.0016508280532434583, + 0.0019241379341110587, + 0.002987499814480543, + 0.0019358124118298292, + 0.013552321121096611, + 0.012788964435458183, + 0.02071228250861168, + 0.09022878110408783, + 0.005748684052377939, + 0.01125410757958889, + 0.01880745403468609, + 0.08294273167848587, + 0.006465406622737646, + 0.005927797872573137, + 0.01084697712212801, + 0.045181237161159515, + 0.003420737572014332, + 0.005395073909312487, + 0.010029747150838375, + 0.042707789689302444, + 0.005446956027299166, + 0.006381672341376543, + 0.008325869217514992, + 0.02614375576376915, + 0.0034066024236381054, + 0.004554540850222111, + 0.00824049860239029, + 0.028186073526740074, + 0.003720998764038086, + 0.0037643916439265013, + 0.005227415822446346, + 0.0037065341603010893, + 0.0022898924071341753, + 0.0029075241181999445, + 0.0053975689224898815, + 0.004103760235011578, + 0.003895045490935445, + 0.004310649819672108, + 0.006158017087727785, + 0.020813237875699997, + 0.0019343869062140584, + 0.003239360637962818, + 0.0052545215003192425, + 0.022308284416794777, + 0.003955426625907421, + 0.0038008398842066526, + 0.006433645263314247, + 0.031668853014707565, + 0.0020810163114219904, + 0.0034321018029004335, + 0.006392746698111296, + 0.029001355171203613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.038765400648117065 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020630886778235435 + ], + "opposite_logprob_tokens": [ + -3.8956308364868164 + ], + "logprob_reward": -0.020630886778235435, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020630886778235435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12487947940826416, + 0.13987435400485992, + 0.05017397180199623, + 0.10268006473779678, + 0.4090125858783722, + 0.02265181764960289, + 0.029265061020851135, + 0.03907181695103645, + 0.028534311801195145, + 0.053856510668992996 + ], + "credit_map": [ + 0.024129623547196388, + 0.014837801456451416, + 0.020208530128002167, + 0.013253787532448769, + 0.009960925206542015, + 0.013457082211971283, + 0.015241982415318489, + 0.01378975436091423, + 0.009363114833831787, + 0.008144143968820572, + 0.023985283449292183, + 0.030522126704454422, + 0.008100271224975586, + 0.008180945180356503, + 0.03322205692529678, + 0.018356436863541603, + 0.005546811502426863, + 0.005722619127482176, + 0.007881727069616318, + 0.005975319538265467, + 0.0036682388745248318, + 0.005635801237076521, + 0.010436550714075565, + 0.005306909326463938, + 0.009493550285696983, + 0.010271679610013962, + 0.01529280748218298, + 0.010493934154510498, + 0.008998888544738293, + 0.010724641382694244, + 0.0243370458483696, + 0.013067521154880524, + 0.020634133368730545, + 0.020458120852708817, + 0.032813649624586105, + 0.14076702296733856, + 0.014015089720487595, + 0.018420899286866188, + 0.028712453320622444, + 0.13319124281406403, + 0.002615328412503004, + 0.0019177814247086644, + 0.002345390385016799, + 0.005307888146489859, + 0.0014002193929627538, + 0.0015076292911544442, + 0.002409749198704958, + 0.005147834308445454, + 0.0025486161466687918, + 0.002841109177097678, + 0.003524750703945756, + 0.00613983953371644, + 0.0018711051670834422, + 0.002013334073126316, + 0.003441530978307128, + 0.006884778384119272, + 0.00258967955596745, + 0.002098057884722948, + 0.003391966223716736, + 0.011861748993396759, + 0.001453503966331482, + 0.0017372100846841931, + 0.0037329162005335093, + 0.01220674067735672, + 0.001589702209457755, + 0.0015480724396184087, + 0.0019291989738121629, + 0.009389766491949558, + 0.0009121937910094857, + 0.0012645843671634793, + 0.0026074827183037996, + 0.009293314069509506, + 0.002523541683331132, + 0.0027535485569387674, + 0.003998237662017345, + 0.01829126849770546, + 0.0016667332965880632, + 0.0019791957456618547, + 0.005117637105286121, + 0.017526356503367424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020630886778235435 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866567999124527 + ], + "opposite_logprob_tokens": [ + -3.048665761947632 + ], + "logprob_reward": -0.04866567999124527, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866567999124527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12746171653270721, + 0.1067584827542305, + 0.029830174520611763, + 0.08115331083536148, + 0.3361188769340515, + 0.032843250781297684, + 0.03357316553592682, + 0.0994851216673851, + 0.0729755163192749, + 0.07980035990476608 + ], + "credit_map": [ + 0.02453118935227394, + 0.013621551916003227, + 0.02381279319524765, + 0.0140821048989892, + 0.007757498882710934, + 0.01191618200391531, + 0.012469310313463211, + 0.01927107572555542, + 0.009243123233318329, + 0.008886462077498436, + 0.021384699270129204, + 0.01915822923183441, + 0.00563589483499527, + 0.00602387823164463, + 0.01936180330812931, + 0.017064400017261505, + 0.003496363293379545, + 0.002688745269551873, + 0.005865562241524458, + 0.004260670393705368, + 0.0019171301973983645, + 0.003064952790737152, + 0.004934266209602356, + 0.003602484939619899, + 0.00980476476252079, + 0.008076193742454052, + 0.012501273304224014, + 0.009109958074986935, + 0.005067359656095505, + 0.008440004661679268, + 0.016548005864024162, + 0.011605752632021904, + 0.016656052321195602, + 0.01692301407456398, + 0.02749170921742916, + 0.11584974825382233, + 0.011799425818026066, + 0.014894536696374416, + 0.023489467799663544, + 0.10901492834091187, + 0.0038287832867354155, + 0.0029975748620927334, + 0.00408233841881156, + 0.0067863548174500465, + 0.001630809041671455, + 0.002179697621613741, + 0.004750449676066637, + 0.006587243638932705, + 0.0047545526176691055, + 0.003336111316457391, + 0.005251400172710419, + 0.005667971912771463, + 0.002073485404253006, + 0.002277243882417679, + 0.0038091158494353294, + 0.0064032841473817825, + 0.0046216887421905994, + 0.003988966811448336, + 0.006269538309425116, + 0.03528030216693878, + 0.0018521299352869391, + 0.003284664824604988, + 0.007367306388914585, + 0.03682052344083786, + 0.0036709890700876713, + 0.003008410567417741, + 0.004784309770911932, + 0.026065871119499207, + 0.0014586752513423562, + 0.0027128385845571756, + 0.0052881804294884205, + 0.025986237451434135, + 0.004033891949802637, + 0.003670588368549943, + 0.005930962041020393, + 0.027352938428521156, + 0.002038594800978899, + 0.003074462991207838, + 0.007058096118271351, + 0.02664083056151867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04866567999124527 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558544985949993 + ], + "opposite_logprob_tokens": [ + -4.387558460235596 + ], + "logprob_reward": -0.012558544985949993, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558544985949993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23017024993896484, + 0.10279181599617004, + 0.02201436087489128, + 0.03672020137310028, + 0.21647465229034424, + 0.20354542136192322, + 0.06953015923500061, + 0.03327293321490288, + 0.025452403351664543, + 0.060027774423360825 + ], + "credit_map": [ + 0.032401300966739655, + 0.027340572327375412, + 0.03288242965936661, + 0.027155691757798195, + 0.026520870625972748, + 0.028847986832261086, + 0.028884584084153175, + 0.02613682486116886, + 0.007139286492019892, + 0.012675671838223934, + 0.017812097445130348, + 0.016200978308916092, + 0.006768262013792992, + 0.008362803608179092, + 0.02176491729915142, + 0.012067802250385284, + 0.002814783249050379, + 0.0035024581011384726, + 0.0038189964834600687, + 0.0025402680039405823, + 0.0020936368964612484, + 0.001979588996618986, + 0.0026092585176229477, + 0.002655372954905033, + 0.0047287833876907825, + 0.0043808394111692905, + 0.006149997003376484, + 0.003957611974328756, + 0.002886133501306176, + 0.004111097659915686, + 0.006457996554672718, + 0.004047746304422617, + 0.01182551123201847, + 0.010601550340652466, + 0.017746947705745697, + 0.07592299580574036, + 0.005426939111202955, + 0.009111378341913223, + 0.016431067138910294, + 0.06940827518701553, + 0.00923428125679493, + 0.009025273844599724, + 0.016762668266892433, + 0.07159141451120377, + 0.005806583911180496, + 0.008172469213604927, + 0.016016481444239616, + 0.06693626195192337, + 0.004618352744728327, + 0.005153034348040819, + 0.008083060383796692, + 0.018284903839230537, + 0.0026955746579915285, + 0.003854277776554227, + 0.006327386479824781, + 0.020513569936156273, + 0.004140942823141813, + 0.002893636701628566, + 0.006367615424096584, + 0.003476236015558243, + 0.002044918481260538, + 0.003026273800060153, + 0.00472591957077384, + 0.006597395054996014, + 0.0026389143895357847, + 0.002629513619467616, + 0.004548338707536459, + 0.004136944655328989, + 0.0012711849994957447, + 0.002130781998857856, + 0.0027258587069809437, + 0.0053708674386143684, + 0.004331440664827824, + 0.004410433117300272, + 0.00611418392509222, + 0.018360745161771774, + 0.002961723832413554, + 0.00285682687535882, + 0.0050560287199914455, + 0.01593640074133873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.012558544985949993 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978767640888691 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.02978767640888691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978767640888691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09494668990373611, + 0.1122620552778244, + 0.05846894532442093, + 0.04093661904335022, + 0.4396421015262604, + 0.027129508554935455, + 0.03558768704533577, + 0.13083869218826294, + 0.02428434044122696, + 0.03590342029929161 + ], + "credit_map": [ + 0.014401000924408436, + 0.009764292277395725, + 0.020163031294941902, + 0.010211546905338764, + 0.006760491989552975, + 0.00783380214124918, + 0.012252889573574066, + 0.013559633865952492, + 0.007558968849480152, + 0.011051229201257229, + 0.024992793798446655, + 0.012903111055493355, + 0.00791002158075571, + 0.00757566886022687, + 0.021002089604735374, + 0.019268173724412918, + 0.00505695678293705, + 0.004987782333046198, + 0.011358378455042839, + 0.012531078420579433, + 0.003294330555945635, + 0.004459125455468893, + 0.009812885895371437, + 0.006968403700739145, + 0.005640807561576366, + 0.004556674975901842, + 0.005026863422244787, + 0.004224310163408518, + 0.004146202001720667, + 0.003917169291526079, + 0.00811529066413641, + 0.005309300497174263, + 0.02091754786670208, + 0.02137216366827488, + 0.035398971289396286, + 0.15120138227939606, + 0.014315483160316944, + 0.019948994740843773, + 0.03125074505805969, + 0.1452368050813675, + 0.0018345688004046679, + 0.0024547474458813667, + 0.002665427979081869, + 0.007558472920209169, + 0.0012098728911951184, + 0.0016362990718334913, + 0.0018098889850080013, + 0.007960230112075806, + 0.004056509584188461, + 0.00171012117061764, + 0.003968509845435619, + 0.009968659840524197, + 0.001343555748462677, + 0.0014147950569167733, + 0.00305776484310627, + 0.010067773051559925, + 0.005469443742185831, + 0.005440181586891413, + 0.008369671180844307, + 0.04699316993355751, + 0.0021584206260740757, + 0.004518903326243162, + 0.00924092996865511, + 0.04864797368645668, + 0.0023274437990039587, + 0.001657460117712617, + 0.0023557767271995544, + 0.005641169846057892, + 0.0011648316867649555, + 0.0015931983944028616, + 0.001990815158933401, + 0.0075536444783210754, + 0.00256438460201025, + 0.0024941551964730024, + 0.00296150054782629, + 0.01124288234859705, + 0.0016653694910928607, + 0.0014948524767532945, + 0.0027747496496886015, + 0.010705525055527687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.02978767640888691 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot depicts a simple physics setup in a craft room. Under the warm, focused light of an unseen desk lamp, a smooth, cylindrical wooden dowel begins to roll steadily down a simple ramp made of pine planks. When it reaches the bottom, it collides with a second, identical dowel resting on the worn surface of a wooden workbench. The impact decisively sends the second dowel sliding to the right, while the first deflects and continues its slow roll to the left. The camera remains fixed as both dowels lose momentum and gently come to rest, now separated and still at the bottom of the ramp. The realistic lighting casts soft, long shadows, accentuating the natural grain of the wood and creating an atmosphere of quiet, focused, hands-on experimentation.", + "seed": 17361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_0/index_367_seed_17361_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": -0.03366656480357051, + "vlm_logprob": -0.03366656480357051, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03366656480357051, + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04959133639931679 + ], + "opposite_logprob_tokens": [ + -3.424591302871704 + ], + "logprob_reward": -0.04959133639931679, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04959133639931679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6116145849227905, + 0.17166756093502045, + 0.08361010253429413, + 0.024902446195483208, + 0.02031181938946247, + 0.01437958050519228, + 0.01140487939119339, + 0.018235722556710243, + 0.01694757118821144, + 0.026925768703222275 + ], + "credit_map": [ + 0.03939250856637955, + 0.033768072724342346, + 0.22376009821891785, + 0.016757914796471596, + 0.038104116916656494, + 0.20439809560775757, + 0.02457767352461815, + 0.0308560561388731, + 0.011287522502243519, + 0.014852740801870823, + 0.016115661710500717, + 0.013479809276759624, + 0.02916366420686245, + 0.02184021845459938, + 0.04547266289591789, + 0.019455278292298317, + 0.006423636805266142, + 0.008086035028100014, + 0.009776919148862362, + 0.008197386749088764, + 0.013654803857207298, + 0.014273351989686489, + 0.011489946395158768, + 0.01170800719410181, + 0.0028305714949965477, + 0.002957829274237156, + 0.0034916994627565145, + 0.0020840608049184084, + 0.003376027336344123, + 0.005397866014391184, + 0.0019729272462427616, + 0.002791464328765869, + 0.0025614581536501646, + 0.0018785736756399274, + 0.00244309869594872, + 0.0016719149425625801, + 0.0029160729609429836, + 0.004251024220138788, + 0.0020769606344401836, + 0.002512714359909296, + 0.0019687023013830185, + 0.0013532245066016912, + 0.0018403964350000024, + 0.001355245360173285, + 0.002183461096137762, + 0.002391840796917677, + 0.0012361017288640141, + 0.0020506056025624275, + 0.001384139759466052, + 0.001034038607031107, + 0.0009930934756994247, + 0.0009829867631196976, + 0.001996999140828848, + 0.0022825137712061405, + 0.0013465702068060637, + 0.0013845370849594474, + 0.002232690341770649, + 0.001361177652142942, + 0.0021846562158316374, + 0.001572036650031805, + 0.0026267708744853735, + 0.0028885765932500362, + 0.0022697639651596546, + 0.003100048052147031, + 0.0019323814194649458, + 0.001422362169250846, + 0.002094145631417632, + 0.0013666017912328243, + 0.0023906840942800045, + 0.0037872998509556055, + 0.0017405337421223521, + 0.002213560277596116, + 0.0026601308491081, + 0.002568380907177925, + 0.0034404455218464136, + 0.0023847606498748064, + 0.004290515091270208, + 0.004046961199492216, + 0.0042632524855434895, + 0.0032713196706026793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04959133639931679 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333815209567547 + ], + "opposite_logprob_tokens": [ + -3.7733380794525146 + ], + "logprob_reward": -0.02333815209567547, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333815209567547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11219716817140579, + 0.6994236707687378, + 0.06415563076734543, + 0.020701901987195015, + 0.010354247875511646, + 0.013157816603779793, + 0.014933456666767597, + 0.009031770750880241, + 0.0123292813077569, + 0.043715063482522964 + ], + "credit_map": [ + 0.00792416650801897, + 0.01838773488998413, + 0.023216091096401215, + 0.007913409732282162, + 0.012150588445365429, + 0.019529877230525017, + 0.014012116938829422, + 0.009063173085451126, + 0.04484788328409195, + 0.05511990562081337, + 0.06556293368339539, + 0.05395809933543205, + 0.10169921070337296, + 0.08647307008504868, + 0.21561473608016968, + 0.07614780217409134, + 0.005899972282350063, + 0.004928445443511009, + 0.006846828386187553, + 0.0065133217722177505, + 0.006951692048460245, + 0.007901661098003387, + 0.013889654539525509, + 0.011224044486880302, + 0.0023275192361325026, + 0.0020664131734520197, + 0.0033772746101021767, + 0.0017271980177611113, + 0.003499493468552828, + 0.0029549046885222197, + 0.002133685629814863, + 0.002615411300212145, + 0.0009867001790553331, + 0.0009351407061330974, + 0.001352344872429967, + 0.0008939729304984212, + 0.0016184381674975157, + 0.001534877228550613, + 0.0013650967739522457, + 0.001667676493525505, + 0.0011080228723585606, + 0.0012886857148259878, + 0.001285974751226604, + 0.0010120088700205088, + 0.001738620805554092, + 0.002925157779827714, + 0.0012930106604471803, + 0.0025063338689506054, + 0.0019004743080586195, + 0.0011782855726778507, + 0.0016398284351453185, + 0.0015619676560163498, + 0.0021890047937631607, + 0.002414629328995943, + 0.002370515139773488, + 0.0016787488711997867, + 0.0010535051114857197, + 0.0008937634411267936, + 0.0011026108404621482, + 0.0009187776013277471, + 0.0016850146930664778, + 0.0012590332189574838, + 0.0011709049576893449, + 0.0009481603046879172, + 0.0016156407073140144, + 0.0012586712837219238, + 0.001538957585580647, + 0.0012156892335042357, + 0.002055202377960086, + 0.001700675580650568, + 0.0014549139887094498, + 0.0014895291533321142, + 0.003027702681720257, + 0.004194516688585281, + 0.004000239074230194, + 0.003337132977321744, + 0.008850608952343464, + 0.007192724384367466, + 0.006607572082430124, + 0.006504560820758343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02333815209567547 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965581744909286 + ], + "opposite_logprob_tokens": [ + -2.9299654960632324 + ], + "logprob_reward": -0.054965581744909286, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965581744909286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18981437385082245, + 0.45532935857772827, + 0.08134881407022476, + 0.04827360063791275, + 0.06493208557367325, + 0.019275587052106857, + 0.03561463952064514, + 0.02025756612420082, + 0.03331821784377098, + 0.05183582007884979 + ], + "credit_map": [ + 0.017267495393753052, + 0.04606309160590172, + 0.025126682594418526, + 0.012442752718925476, + 0.020795904099941254, + 0.03525731712579727, + 0.01520971767604351, + 0.01765141636133194, + 0.030734756961464882, + 0.03772531822323799, + 0.04144010320305824, + 0.0336066372692585, + 0.047597821801900864, + 0.0678769201040268, + 0.13381987810134888, + 0.06252788752317429, + 0.010923963971436024, + 0.006404203828424215, + 0.008117755874991417, + 0.006351396907120943, + 0.009559483267366886, + 0.012609289027750492, + 0.01490646880120039, + 0.012476254254579544, + 0.0042394837364554405, + 0.006843486335128546, + 0.004492310807108879, + 0.002916802652180195, + 0.008601938374340534, + 0.007429369725286961, + 0.006150313653051853, + 0.007599893491715193, + 0.0076834214851260185, + 0.005629038438200951, + 0.009030697867274284, + 0.006088728550821543, + 0.009763051755726337, + 0.011790758930146694, + 0.0071440753526985645, + 0.0078023155219852924, + 0.0021929387003183365, + 0.001793618081137538, + 0.002633147407323122, + 0.0016072290018200874, + 0.0031758935656398535, + 0.0029893030878156424, + 0.0020218181889504194, + 0.002861639019101858, + 0.004406733904033899, + 0.0032298010773956776, + 0.003730000462383032, + 0.0034100410994142294, + 0.0059113516472280025, + 0.005776127800345421, + 0.005459983367472887, + 0.003690598765388131, + 0.0032833325676620007, + 0.0018174578435719013, + 0.0018769894959405065, + 0.0019643858540803194, + 0.0021586178336292505, + 0.004550964571535587, + 0.0021643361542373896, + 0.002441482385620475, + 0.003089067991822958, + 0.0029933892656117678, + 0.0036109599750488997, + 0.003175202291458845, + 0.005198219791054726, + 0.005976083688437939, + 0.004460357129573822, + 0.004814939107745886, + 0.004221583250910044, + 0.004627615679055452, + 0.005815425422042608, + 0.004168722312897444, + 0.008674168959259987, + 0.009275810793042183, + 0.007658418733626604, + 0.0073940749280154705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.054965581744909286 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006749450694769621 + ], + "opposite_logprob_tokens": [ + -5.006749629974365 + ], + "logprob_reward": -0.006749450694769621, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749450694769621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21844545006752014, + 0.22444632649421692, + 0.128896102309227, + 0.07220473885536194, + 0.07703036069869995, + 0.05357367545366287, + 0.04267944023013115, + 0.06909482926130295, + 0.06570899486541748, + 0.047920044511556625 + ], + "credit_map": [ + 0.016401391476392746, + 0.04374166205525398, + 0.029115956276655197, + 0.015758676454424858, + 0.02206641063094139, + 0.04664799943566322, + 0.01992231048643589, + 0.02479102835059166, + 0.017279379069805145, + 0.020459448918700218, + 0.024082235991954803, + 0.019197562709450722, + 0.032813865691423416, + 0.05280196666717529, + 0.03376820683479309, + 0.02404366061091423, + 0.012415335513651371, + 0.01168469525873661, + 0.014795380644500256, + 0.01532441470772028, + 0.014065202325582504, + 0.012591253966093063, + 0.0273694209754467, + 0.020650403574109077, + 0.006946748588234186, + 0.008439428173005581, + 0.006449950393289328, + 0.005877386312931776, + 0.014861732721328735, + 0.013698313385248184, + 0.0074167679995298386, + 0.008514410816133022, + 0.00906202383339405, + 0.007148353382945061, + 0.009901270270347595, + 0.007541613653302193, + 0.010786643251776695, + 0.013754432089626789, + 0.007600785698741674, + 0.011235239915549755, + 0.00457646232098341, + 0.00563926761969924, + 0.006327121052891016, + 0.004366672597825527, + 0.006283426657319069, + 0.013685299083590508, + 0.004660515580326319, + 0.008034912869334221, + 0.0035004708915948868, + 0.004451077897101641, + 0.005178108811378479, + 0.002960802521556616, + 0.007966367527842522, + 0.010120447725057602, + 0.004933866672217846, + 0.0035682993475347757, + 0.008824246004223824, + 0.007500567939132452, + 0.007287552114576101, + 0.007445215247571468, + 0.009867438115179539, + 0.010273769497871399, + 0.007610275410115719, + 0.010285759344696999, + 0.00895535759627819, + 0.005983928684145212, + 0.0070922295562922955, + 0.005475250072777271, + 0.008420972153544426, + 0.01555855292826891, + 0.007644947152584791, + 0.006577757187187672, + 0.004446099046617746, + 0.004104065243154764, + 0.005546309519559145, + 0.003903142875060439, + 0.009205684065818787, + 0.008932930417358875, + 0.00640213955193758, + 0.005379676818847656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006749450694769621 + }, + { + "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 craft room with a wooden workbench and a ramp made of pine planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a smooth cylindrical wooden dowel roll steadily down the ramp and collide with a second identical dowel at the bottom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"After the collision, does the first dowel deflect and roll slowly to the left while the second dowel slides to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lighting come from an unseen desk lamp, casting soft long shadows and highlighting the natural grain of the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a static, slightly high-angled medium shot throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368830308318138 + ], + "opposite_logprob_tokens": [ + -3.4086883068084717 + ], + "logprob_reward": -0.03368830308318138, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368830308318138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5717716217041016, + 0.1566212773323059, + 0.11357540637254715, + 0.015432707965373993, + 0.02830112725496292, + 0.018367359414696693, + 0.01699487864971161, + 0.027285005897283554, + 0.028648026287555695, + 0.02300257235765457 + ], + "credit_map": [ + 0.03655238449573517, + 0.025003330782055855, + 0.2218036949634552, + 0.014948052354156971, + 0.03331414982676506, + 0.19136640429496765, + 0.020768195390701294, + 0.02801540121436119, + 0.019267426803708076, + 0.012744826264679432, + 0.01661568321287632, + 0.014878581278026104, + 0.016283312812447548, + 0.03281756117939949, + 0.024814946576952934, + 0.019198941066861153, + 0.01480226218700409, + 0.011049230583012104, + 0.010599764995276928, + 0.009733396582305431, + 0.015440086834132671, + 0.01266709715127945, + 0.02267010696232319, + 0.016613462939858437, + 0.002492147497832775, + 0.001425696536898613, + 0.0015703558456152678, + 0.001259456155821681, + 0.0031335405074059963, + 0.0016989713767543435, + 0.002202205825597048, + 0.0016503348015248775, + 0.003525015665218234, + 0.0023863068781793118, + 0.0036673275753855705, + 0.0025622660759836435, + 0.0046449811197817326, + 0.0051323845982551575, + 0.0027729168068617582, + 0.003609929932281375, + 0.001684721326455474, + 0.0018270566361024976, + 0.0023923972621560097, + 0.001405098824761808, + 0.0028497406747192144, + 0.004373225849121809, + 0.0017188285710290074, + 0.002116290619596839, + 0.0013440027832984924, + 0.001781285391189158, + 0.001838476164266467, + 0.0013954580063000321, + 0.002869123127311468, + 0.0032734768465161324, + 0.0026058065705001354, + 0.0018872495274990797, + 0.003343253629282117, + 0.0024479448329657316, + 0.002849074313417077, + 0.002359780017286539, + 0.004147949628531933, + 0.005064403638243675, + 0.003633366897702217, + 0.003439231775701046, + 0.004090782254934311, + 0.002783441450446844, + 0.0030509757343679667, + 0.0027554486878216267, + 0.004421544726938009, + 0.0048099784180521965, + 0.003675131592899561, + 0.0030607248190790415, + 0.0013814409030601382, + 0.0021460331045091152, + 0.0023621092550456524, + 0.001795260701328516, + 0.004703988786786795, + 0.004386363085359335, + 0.0028823537286370993, + 0.003345021279528737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03368830308318138 + } + ], + "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/0000017/group_1/index_tensor([5510, 5510])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_tensor([5510, 5510])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a02304249da8ef4395a9582660570685c5e0de1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_tensor([5510, 5510])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "documentary-style video, shot from a steady, handheld medium shot, held slightly upward, slowly and almost imperceptibly approaches its subject, focusing on an elderly explorer seated around a campfire. His face is weathered, deeply wrinkled like a dry riverbed, his bronzed skin gleaming oily in the dancing flames. An old scar runs from his left eyebrow to his temple. His graying hair is a little unkempt, and his thick beard quivers slightly with the rhythm of his speech. He wears a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest, frayed at the edges. He sits on a rough log, leaning slightly forward, his eyes sharp and focused, recounting past events to his companions across the fire. Their figures are blurred in the shot, only their profiles, illuminated by the firelight, their focused faces visible. The veteran explorer doesn't simply narrate silently; his movements are imbued with power: sometimes he uses his calloused right hand to gesture in the air to depict a treacherous terrain, sometimes he clenches his fist, as if reliving the tense moments of the past. Each gesture is precise and expressive. The primary light source for the entire scene comes from the blazing campfire in the center. The warm orange-red glow casts deep, shifting shadows on his face, separating him from the deep, cold coniferous forest nightscape and the star-studded sky behind him. This creates an atmosphere of both intense and captivating intimacy, as if the audience has become one of the listeners to this evening's conversation.", + "seed": 29843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29843_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": -0.022374965995550156, + "vlm_logprob": -0.022374965995550156, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.022374965995550156, + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05186745151877403 + ], + "opposite_logprob_tokens": [ + -3.0518674850463867 + ], + "logprob_reward": -0.05186745151877403, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05186745151877403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09329749643802643, + 0.03546235337853432, + 0.06430947780609131, + 0.18711590766906738, + 0.050213638693094254, + 0.08594350516796112, + 0.059394191950559616, + 0.04473511502146721, + 0.3561680316925049, + 0.023360248655080795 + ], + "credit_map": [ + 0.012368468567728996, + 0.013053646311163902, + 0.018135495483875275, + 0.012435401789844036, + 0.008290847763419151, + 0.009757746011018753, + 0.009238156490027905, + 0.010017734952270985, + 0.004543769638985395, + 0.0060841417871415615, + 0.005717553198337555, + 0.004434662871062756, + 0.004524798132479191, + 0.0034412394743412733, + 0.0037821924779564142, + 0.002933995798230171, + 0.005828707478940487, + 0.010081954300403595, + 0.010232717730104923, + 0.004835038911551237, + 0.00902180653065443, + 0.0072200000286102295, + 0.004606746602803469, + 0.012482506223022938, + 0.0185859352350235, + 0.00959849450737238, + 0.061981841921806335, + 0.011626374907791615, + 0.015159056521952152, + 0.05452689900994301, + 0.00689076678827405, + 0.008746547624468803, + 0.007842461578547955, + 0.008351679891347885, + 0.007609296590089798, + 0.004848852753639221, + 0.007339168805629015, + 0.006416564341634512, + 0.003862574929371476, + 0.003943036310374737, + 0.007795317098498344, + 0.012661771848797798, + 0.014113085344433784, + 0.010741283185780048, + 0.012353349477052689, + 0.014408816583454609, + 0.006074819713830948, + 0.007795062381774187, + 0.008798915892839432, + 0.005211396608501673, + 0.0061766114085912704, + 0.007723737042397261, + 0.011892656795680523, + 0.006818636320531368, + 0.00615245895460248, + 0.006619777996093035, + 0.0075117070227861404, + 0.006233889609575272, + 0.004730921238660812, + 0.005540878511965275, + 0.005514855496585369, + 0.005990243516862392, + 0.0043396069668233395, + 0.0048730140551924706, + 0.03788258880376816, + 0.0593826100230217, + 0.03710339590907097, + 0.03832505643367767, + 0.058097340166568756, + 0.06623321771621704, + 0.02655980736017227, + 0.03258403390645981, + 0.002144221216440201, + 0.0034253010526299477, + 0.005464829504489899, + 0.003076861845329404, + 0.002035189187154174, + 0.003131435951218009, + 0.0015750336460769176, + 0.0025073769502341747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.05186745151877403 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012569492682814598 + ], + "opposite_logprob_tokens": [ + -4.387569427490234 + ], + "logprob_reward": -0.012569492682814598, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012569492682814598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10515782237052917, + 0.10003499686717987, + 0.023040585219860077, + 0.15799368917942047, + 0.06610395014286041, + 0.057392578572034836, + 0.048862915486097336, + 0.0656086876988411, + 0.09314748644828796, + 0.28265732526779175 + ], + "credit_map": [ + 0.012229733169078827, + 0.01684403046965599, + 0.01869618147611618, + 0.016086138784885406, + 0.008450607769191265, + 0.00908528733998537, + 0.011207172647118568, + 0.012558657675981522, + 0.009511248208582401, + 0.021208571270108223, + 0.01271012518554926, + 0.012973365373909473, + 0.015723425894975662, + 0.009918750263750553, + 0.009613488800823689, + 0.008376027457416058, + 0.0021570243407040834, + 0.0031785708852112293, + 0.0031957917381078005, + 0.0024962248280644417, + 0.00299263303168118, + 0.0038429489359259605, + 0.0024981615133583546, + 0.002679226454347372, + 0.016803475096821785, + 0.011179355904459953, + 0.04593750834465027, + 0.010544496588408947, + 0.01676284894347191, + 0.04164521396160126, + 0.006807082798331976, + 0.00831370148807764, + 0.00730378832668066, + 0.00841411855071783, + 0.012867849320173264, + 0.007557795383036137, + 0.008680331520736217, + 0.010989471338689327, + 0.005572039633989334, + 0.004718551877886057, + 0.005752294324338436, + 0.011438525281846523, + 0.00595848448574543, + 0.007343794219195843, + 0.008014122024178505, + 0.006914607714861631, + 0.005652599036693573, + 0.006318142171949148, + 0.006217929534614086, + 0.005050864536315203, + 0.0045221056789159775, + 0.005832881201058626, + 0.0103157302364707, + 0.006723507307469845, + 0.004925436805933714, + 0.005274457857012749, + 0.011638971045613289, + 0.01181040983647108, + 0.009772507473826408, + 0.007537975907325745, + 0.006391528062522411, + 0.007003611419349909, + 0.005081202369183302, + 0.006372481118887663, + 0.00985135231167078, + 0.015786219388246536, + 0.008647016249597073, + 0.010740338824689388, + 0.013156854547560215, + 0.017224866896867752, + 0.007064792327582836, + 0.010676045902073383, + 0.02071595937013626, + 0.03322271257638931, + 0.11158785969018936, + 0.030943067744374275, + 0.011080658063292503, + 0.04147078096866608, + 0.010789343155920506, + 0.022846927866339684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.012569492682814598 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008665459230542183 + ], + "opposite_logprob_tokens": [ + -4.758665561676025 + ], + "logprob_reward": -0.008665459230542183, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008665459230542183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05950135737657547, + 0.025940371677279472, + 0.04428049176931381, + 0.1149359941482544, + 0.07003681361675262, + 0.038450937718153, + 0.026305066421628, + 0.07180886715650558, + 0.14148572087287903, + 0.4072543978691101 + ], + "credit_map": [ + 0.005893167573958635, + 0.01024585124105215, + 0.009891154244542122, + 0.008482467383146286, + 0.00532666128128767, + 0.005609283223748207, + 0.006169009022414684, + 0.007883764803409576, + 0.0022531538270413876, + 0.004733624868094921, + 0.0034154588356614113, + 0.002925009001046419, + 0.0034701251424849033, + 0.004080979619175196, + 0.0027509413193911314, + 0.0023110785987228155, + 0.004887303337454796, + 0.00844633113592863, + 0.007080161012709141, + 0.0037774527445435524, + 0.005355697590857744, + 0.0071908943355083466, + 0.002967143664136529, + 0.004575506318360567, + 0.013300225138664246, + 0.006995648145675659, + 0.038128964602947235, + 0.006853797473013401, + 0.006214702036231756, + 0.03381095826625824, + 0.0044571831822395325, + 0.00517450924962759, + 0.00639285147190094, + 0.006758465897291899, + 0.016728390008211136, + 0.007238015998154879, + 0.007820205762982368, + 0.014492674730718136, + 0.00561866769567132, + 0.004987543914467096, + 0.005772271193563938, + 0.006572007201611996, + 0.004102811682969332, + 0.00521629024296999, + 0.00382111850194633, + 0.004634065553545952, + 0.0034312757197767496, + 0.004901098553091288, + 0.0031363496091216803, + 0.003577343188226223, + 0.0025576672051101923, + 0.0032320942264050245, + 0.0037728233728557825, + 0.003570658154785633, + 0.003381887450814247, + 0.0030762432143092155, + 0.006825210992246866, + 0.006838304456323385, + 0.013401119969785213, + 0.014624911360442638, + 0.006828484125435352, + 0.008121778257191181, + 0.0077422321774065495, + 0.007426820695400238, + 0.012945887632668018, + 0.020640725269913673, + 0.012522059492766857, + 0.015405762940645218, + 0.014965892769396305, + 0.022769788280129433, + 0.010620605200529099, + 0.0316149964928627, + 0.029587360098958015, + 0.048351988196372986, + 0.16110551357269287, + 0.0440821647644043, + 0.01583772525191307, + 0.061695411801338196, + 0.014598536305129528, + 0.03199569135904312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.008665459230542183 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15930496156215668, + 0.07456650584936142, + 0.04891311377286911, + 0.0953996330499649, + 0.13348104059696198, + 0.07175847142934799, + 0.07328956574201584, + 0.15304885804653168, + 0.11249149590730667, + 0.07774632424116135 + ], + "credit_map": [ + 0.019882624968886375, + 0.026330364868044853, + 0.02838308736681938, + 0.0197627954185009, + 0.01726694218814373, + 0.015877921134233475, + 0.015897782519459724, + 0.01590343751013279, + 0.008151471614837646, + 0.013234197162091732, + 0.010406232438981533, + 0.010153871029615402, + 0.01126438844949007, + 0.007536634337157011, + 0.0074903215281665325, + 0.006329391151666641, + 0.004306787624955177, + 0.0063659814186394215, + 0.008235674351453781, + 0.005294904112815857, + 0.006847522221505642, + 0.006882893852889538, + 0.004113179165869951, + 0.006866168696433306, + 0.0106271393597126, + 0.00619155541062355, + 0.026672083884477615, + 0.006589957512915134, + 0.008594203740358353, + 0.024904966354370117, + 0.004439990036189556, + 0.007379736751317978, + 0.020330460742115974, + 0.023003045469522476, + 0.01398595329374075, + 0.01438823901116848, + 0.01730497181415558, + 0.016471294686198235, + 0.011455432511866093, + 0.016541633754968643, + 0.0059286197647452354, + 0.009639027528464794, + 0.015387538820505142, + 0.0070072365924716, + 0.00903900247067213, + 0.013722318224608898, + 0.004707529675215483, + 0.006327195558696985, + 0.00807665754109621, + 0.010737157426774502, + 0.009521730244159698, + 0.011140497401356697, + 0.010421764105558395, + 0.008742078207433224, + 0.007406693417578936, + 0.007242986932396889, + 0.0187684353441, + 0.0193344596773386, + 0.01804710365831852, + 0.031834449619054794, + 0.01626535877585411, + 0.023223647847771645, + 0.015301176346838474, + 0.010274223983287811, + 0.009082973003387451, + 0.016772976145148277, + 0.017604177817702293, + 0.012289421632885933, + 0.016618015244603157, + 0.02071692794561386, + 0.007984084077179432, + 0.011422915384173393, + 0.0068180495873093605, + 0.010692963376641273, + 0.027722172439098358, + 0.008113615214824677, + 0.003996743354946375, + 0.010447829961776733, + 0.0034462600015103817, + 0.006508681923151016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01819257251918316 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10747677087783813, + 0.08565269410610199, + 0.08981327712535858, + 0.15945802628993988, + 0.04741353169083595, + 0.054086796939373016, + 0.08837713301181793, + 0.06793200224637985, + 0.20289549231529236, + 0.09689421206712723 + ], + "credit_map": [ + 0.009766396135091782, + 0.016408920288085938, + 0.01591714657843113, + 0.017642822116613388, + 0.007978704757988453, + 0.010781671851873398, + 0.014704028144478798, + 0.01427709124982357, + 0.011260519735515118, + 0.017171965911984444, + 0.014469758607447147, + 0.00842319056391716, + 0.01319439709186554, + 0.007768373470753431, + 0.007299778051674366, + 0.006064714398235083, + 0.010050185956060886, + 0.01666487380862236, + 0.013035298325121403, + 0.009378110989928246, + 0.010250107385218143, + 0.013543549925088882, + 0.007838922552764416, + 0.009052237495779991, + 0.016595015302300453, + 0.006979295983910561, + 0.051915496587753296, + 0.01069395150989294, + 0.01070677675306797, + 0.04674173519015312, + 0.0066565945744514465, + 0.009169178083539009, + 0.0063567692413926125, + 0.00701001426205039, + 0.00797371007502079, + 0.004140030592679977, + 0.006512091029435396, + 0.006867627613246441, + 0.00365427671931684, + 0.004899018444120884, + 0.0067423558793962, + 0.009666423313319683, + 0.007253884803503752, + 0.005481585394591093, + 0.006530091166496277, + 0.006885536480695009, + 0.004527897574007511, + 0.006999025121331215, + 0.012056935578584671, + 0.009338784962892532, + 0.009567643515765667, + 0.010405505076050758, + 0.017595836892724037, + 0.011236552149057388, + 0.00870707631111145, + 0.0094688069075346, + 0.0073142084293067455, + 0.01187086757272482, + 0.008132165297865868, + 0.011502803303301334, + 0.0076682232320308685, + 0.010252374224364758, + 0.006095256190747023, + 0.005096112843602896, + 0.019807154312729836, + 0.04120682179927826, + 0.018547838553786278, + 0.022109834477305412, + 0.028728261590003967, + 0.03314175084233284, + 0.017822589725255966, + 0.021531255915760994, + 0.008230955339968204, + 0.012550289742648602, + 0.03567534685134888, + 0.011635693721473217, + 0.0049066138453781605, + 0.006930723320692778, + 0.003942694514989853, + 0.013021900318562984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.020579854026436806 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style video, shot from a steady, handheld medium shot, held slightly upward, slowly and almost imperceptibly approaches its subject, focusing on an elderly explorer seated around a campfire. His face is weathered, deeply wrinkled like a dry riverbed, his bronzed skin gleaming oily in the dancing flames. An old scar runs from his left eyebrow to his temple. His graying hair is a little unkempt, and his thick beard quivers slightly with the rhythm of his speech. He wears a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest, frayed at the edges. He sits on a rough log, leaning slightly forward, his eyes sharp and focused, recounting past events to his companions across the fire. Their figures are blurred in the shot, only their profiles, illuminated by the firelight, their focused faces visible. The veteran explorer doesn't simply narrate silently; his movements are imbued with power: sometimes he uses his calloused right hand to gesture in the air to depict a treacherous terrain, sometimes he clenches his fist, as if reliving the tense moments of the past. Each gesture is precise and expressive. The primary light source for the entire scene comes from the blazing campfire in the center. The warm orange-red glow casts deep, shifting shadows on his face, separating him from the deep, cold coniferous forest nightscape and the star-studded sky behind him. This creates an atmosphere of both intense and captivating intimacy, as if the audience has become one of the listeners to this evening's conversation.", + "seed": 29844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29844_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": -0.021742330864071846, + "vlm_logprob": -0.021742330864071846, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.021742330864071846, + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05328848212957382 + ], + "opposite_logprob_tokens": [ + -3.053288459777832 + ], + "logprob_reward": -0.05328848212957382, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05328848212957382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06774585694074631, + 0.271811842918396, + 0.05631528049707413, + 0.03993811458349228, + 0.1488061249256134, + 0.1468726247549057, + 0.05853689834475517, + 0.041090287268161774, + 0.024124735966324806, + 0.14475831389427185 + ], + "credit_map": [ + 0.008035409264266491, + 0.010686088353395462, + 0.009849210269749165, + 0.0075287423096597195, + 0.007953023537993431, + 0.008146719075739384, + 0.009608861058950424, + 0.005937803070992231, + 0.016794275492429733, + 0.016077185049653053, + 0.07849234342575073, + 0.018590303137898445, + 0.026141555979847908, + 0.09488499909639359, + 0.011643176898360252, + 0.009187989868223667, + 0.004734870512038469, + 0.011406919918954372, + 0.007027484476566315, + 0.006890407297760248, + 0.006355171091854572, + 0.007900367490947247, + 0.005320923402905464, + 0.006679135840386152, + 0.002982890233397484, + 0.007573397364467382, + 0.004632961470633745, + 0.005519649479538202, + 0.004453680012375116, + 0.005523888394236565, + 0.005589619278907776, + 0.0036620309110730886, + 0.01620352268218994, + 0.027658425271511078, + 0.016714831814169884, + 0.01969933696091175, + 0.018197260797023773, + 0.01992209255695343, + 0.0115515673533082, + 0.01885908842086792, + 0.011733020655810833, + 0.0351901575922966, + 0.02039693295955658, + 0.021044092252850533, + 0.013491629622876644, + 0.020914701744914055, + 0.01501765288412571, + 0.009084431454539299, + 0.005030475091189146, + 0.011265615932643414, + 0.0063249957747757435, + 0.014882753603160381, + 0.003197798738256097, + 0.005588441621512175, + 0.005958189722150564, + 0.006288623437285423, + 0.0032985720317810774, + 0.005570131819695234, + 0.004252403974533081, + 0.007444742135703564, + 0.003964717965573072, + 0.00636551296338439, + 0.004542519338428974, + 0.005651686806231737, + 0.002143088262528181, + 0.003390745259821415, + 0.0033011690247803926, + 0.004719369579106569, + 0.0026629024650901556, + 0.0027135780546814203, + 0.0031866810750216246, + 0.002007201546803117, + 0.0107578719034791, + 0.03518596664071083, + 0.014052280224859715, + 0.02627703920006752, + 0.011068557389080524, + 0.0188872292637825, + 0.018962612375617027, + 0.009566744789481163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.05328848212957382 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008670658804476261 + ], + "opposite_logprob_tokens": [ + -4.758670806884766 + ], + "logprob_reward": -0.008670658804476261, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008670658804476261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03145240992307663, + 0.6708801984786987, + 0.02379949390888214, + 0.10923779755830765, + 0.04497823864221573, + 0.03922761231660843, + 0.017100529745221138, + 0.03048035129904747, + 0.010165448300540447, + 0.022677864879369736 + ], + "credit_map": [ + 0.004912341479212046, + 0.005426972638815641, + 0.0035046807024627924, + 0.003040231531485915, + 0.0038425391539931297, + 0.004256990272551775, + 0.003663499839603901, + 0.002805158030241728, + 0.04067298397421837, + 0.0236999299377203, + 0.21181204915046692, + 0.04334283620119095, + 0.05953573063015938, + 0.23779605329036713, + 0.03499207645654678, + 0.019028538838028908, + 0.0022810788359493017, + 0.004766533616930246, + 0.0031329388730227947, + 0.002813389291986823, + 0.0020903386175632477, + 0.003054898465052247, + 0.002448481507599354, + 0.0032118367962539196, + 0.006002643145620823, + 0.022845854982733727, + 0.008749720640480518, + 0.013661976903676987, + 0.014371677301824093, + 0.016490669921040535, + 0.010930424556136131, + 0.016184834763407707, + 0.003745689056813717, + 0.013549631461501122, + 0.003326810896396637, + 0.005780354142189026, + 0.004148406907916069, + 0.005625289399176836, + 0.00426373491063714, + 0.004538326058536768, + 0.003192297415807843, + 0.008823618292808533, + 0.005149753764271736, + 0.005317901726812124, + 0.0032591253984719515, + 0.006174380891025066, + 0.0047968653962016106, + 0.0025136691983789206, + 0.0017786528915166855, + 0.002448220271617174, + 0.002023821696639061, + 0.00403182627633214, + 0.0012900253059342504, + 0.002078794175758958, + 0.0017201157752424479, + 0.0017290746327489614, + 0.00246426067315042, + 0.006778524722903967, + 0.0029564599972218275, + 0.004862635862082243, + 0.00231628748588264, + 0.004545549396425486, + 0.0027550463564693928, + 0.0038015886675566435, + 0.0008946888265199959, + 0.0017329343827441335, + 0.0017280812608078122, + 0.0016019507311284542, + 0.0009794466895982623, + 0.0009672136511653662, + 0.001559313852339983, + 0.0007018200703896582, + 0.0019332674564793706, + 0.0037829510401934385, + 0.002810640959069133, + 0.0031405079644173384, + 0.002206675475463271, + 0.004834136925637722, + 0.002167923143133521, + 0.0018017624970525503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.008670658804476261 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13140830397605896, + 0.10680560767650604, + 0.09896830469369888, + 0.12420766055583954, + 0.06687057763338089, + 0.10358044505119324, + 0.0971880555152893, + 0.08072075992822647, + 0.04449286311864853, + 0.14575742185115814 + ], + "credit_map": [ + 0.014283627271652222, + 0.025252997875213623, + 0.013831240125000477, + 0.015427062287926674, + 0.01824381947517395, + 0.016581613570451736, + 0.01486836839467287, + 0.012919570319354534, + 0.007233666256070137, + 0.012001574039459229, + 0.010973911732435226, + 0.01506360899657011, + 0.014859461225569248, + 0.022968627512454987, + 0.01127934455871582, + 0.012425409629940987, + 0.006922291591763496, + 0.012378446757793427, + 0.0139438072219491, + 0.013413350097835064, + 0.014227778650820255, + 0.014538312330842018, + 0.014008383266627789, + 0.009535935707390308, + 0.007420817390084267, + 0.019223446026444435, + 0.01557563804090023, + 0.018095707520842552, + 0.014083362184464931, + 0.01646861433982849, + 0.018824150785803795, + 0.014515913091599941, + 0.007286328822374344, + 0.01229727454483509, + 0.010215120390057564, + 0.00804899726063013, + 0.006857617758214474, + 0.009576969780027866, + 0.005657355301082134, + 0.006930910050868988, + 0.009393516927957535, + 0.020261388272047043, + 0.011827175505459309, + 0.014050778932869434, + 0.007311430759727955, + 0.023382272571325302, + 0.010506605729460716, + 0.00684726145118475, + 0.008315193466842175, + 0.0167397391051054, + 0.010480881668627262, + 0.02950999140739441, + 0.007054341956973076, + 0.0070504299364984035, + 0.010859711095690727, + 0.0071777584962546825, + 0.007125430274754763, + 0.025236831977963448, + 0.0076352814212441444, + 0.011906063184142113, + 0.006919096689671278, + 0.006608234718441963, + 0.009013740345835686, + 0.0062760766595602036, + 0.003844461403787136, + 0.009075038135051727, + 0.0060280729085206985, + 0.007008746732026339, + 0.005618916358798742, + 0.004453248344361782, + 0.005343542899936438, + 0.003120837965980172, + 0.010607129894196987, + 0.044812947511672974, + 0.014216774143278599, + 0.024197716265916824, + 0.013179076835513115, + 0.013499307446181774, + 0.018114270642399788, + 0.007130196318030357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01821083389222622 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256284959614277 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256284959614277, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256284959614277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0989336371421814, + 0.2177494466304779, + 0.3071102201938629, + 0.05279511213302612, + 0.11038418859243393, + 0.05562791973352432, + 0.05055336654186249, + 0.02924175374209881, + 0.031567275524139404, + 0.04603709280490875 + ], + "credit_map": [ + 0.010856433771550655, + 0.02046504244208336, + 0.012824270874261856, + 0.008972526527941227, + 0.013736167922616005, + 0.010545431636273861, + 0.008156820200383663, + 0.013376944698393345, + 0.014491468667984009, + 0.015102477744221687, + 0.06895481050014496, + 0.01647220179438591, + 0.017414389178156853, + 0.06541676074266434, + 0.010138597339391708, + 0.009758739732205868, + 0.0204904917627573, + 0.05305754393339157, + 0.03965359926223755, + 0.03853374347090721, + 0.037768762558698654, + 0.040631260722875595, + 0.037749942392110825, + 0.03922489285469055, + 0.0030775268096476793, + 0.009608602151274681, + 0.003740591462701559, + 0.006432443857192993, + 0.0062467860989272594, + 0.007214282173663378, + 0.0058166394010186195, + 0.01065823994576931, + 0.008851420134305954, + 0.02290373481810093, + 0.010233473032712936, + 0.01175539381802082, + 0.009043834172189236, + 0.013581349514424801, + 0.011193279176950455, + 0.022821711376309395, + 0.0051641943864524364, + 0.006191322114318609, + 0.005742487031966448, + 0.006871596910059452, + 0.005554783158004284, + 0.013034693896770477, + 0.008478104136884212, + 0.0045907385647296906, + 0.005839658435434103, + 0.012240970507264137, + 0.006100178230553865, + 0.00635237293317914, + 0.004842326510697603, + 0.005095119588077068, + 0.0051014963537454605, + 0.004981242120265961, + 0.00242342590354383, + 0.006098375655710697, + 0.0031725233420729637, + 0.006232514511793852, + 0.001982043031603098, + 0.00348876416683197, + 0.0029188196640461683, + 0.0029252865351736546, + 0.002255597384646535, + 0.005437527317553759, + 0.0049524446949362755, + 0.004747699946165085, + 0.0030143216717988253, + 0.004427128471434116, + 0.004292151890695095, + 0.002440405311062932, + 0.0035186558961868286, + 0.013688704930245876, + 0.004608712624758482, + 0.0072099557146430016, + 0.0027536670677363873, + 0.005659983493387699, + 0.005165550857782364, + 0.003431858029216528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.005256284959614277 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328539453446865 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.02328539453446865, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328539453446865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14482729136943817, + 0.13545915484428406, + 0.10151272267103195, + 0.07924143970012665, + 0.062371332198381424, + 0.08053008466959, + 0.07028071582317352, + 0.12445566803216934, + 0.12069676071405411, + 0.08062489330768585 + ], + "credit_map": [ + 0.020243298262357712, + 0.023062456399202347, + 0.01593962125480175, + 0.014089456759393215, + 0.020262889564037323, + 0.021846871823072433, + 0.015737958252429962, + 0.013644738122820854, + 0.010593762621283531, + 0.01587132178246975, + 0.018891088664531708, + 0.017621444538235664, + 0.014651212841272354, + 0.022637560963630676, + 0.020029090344905853, + 0.015163671225309372, + 0.008585578761994839, + 0.015638442710042, + 0.010648010298609734, + 0.013910296373069286, + 0.012822341173887253, + 0.014574647881090641, + 0.01273841317743063, + 0.01259499043226242, + 0.006235230714082718, + 0.014742877334356308, + 0.009466784074902534, + 0.007782078348100185, + 0.007411593105643988, + 0.010671371594071388, + 0.0066834744065999985, + 0.016248028725385666, + 0.00596959562972188, + 0.013769038021564484, + 0.005638351663947105, + 0.007423154544085264, + 0.007876253686845303, + 0.007351561449468136, + 0.008724245242774487, + 0.0056191314943134785, + 0.008251863531768322, + 0.012878093868494034, + 0.011156859807670116, + 0.01043421495705843, + 0.007986452430486679, + 0.012726321816444397, + 0.012152336537837982, + 0.00494394451379776, + 0.008665812201797962, + 0.0089942067861557, + 0.006954750511795282, + 0.02155361697077751, + 0.004627269692718983, + 0.007729357108473778, + 0.005910868756473064, + 0.005844833794981241, + 0.009971614927053452, + 0.034553900361061096, + 0.010298493318259716, + 0.017938047647476196, + 0.011228718794882298, + 0.015928437933325768, + 0.012915868312120438, + 0.011620587669312954, + 0.009259789250791073, + 0.032338693737983704, + 0.012001143768429756, + 0.01964438520371914, + 0.009357730858027935, + 0.007484814152121544, + 0.014306356199085712, + 0.01630384661257267, + 0.007211124524474144, + 0.02272491715848446, + 0.0077308048494160175, + 0.01335025392472744, + 0.00667891139164567, + 0.007652687840163708, + 0.010671229101717472, + 0.004604965448379517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02328539453446865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style video, shot from a steady, handheld medium shot, held slightly upward, slowly and almost imperceptibly approaches its subject, focusing on an elderly explorer seated around a campfire. His face is weathered, deeply wrinkled like a dry riverbed, his bronzed skin gleaming oily in the dancing flames. An old scar runs from his left eyebrow to his temple. His graying hair is a little unkempt, and his thick beard quivers slightly with the rhythm of his speech. He wears a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest, frayed at the edges. He sits on a rough log, leaning slightly forward, his eyes sharp and focused, recounting past events to his companions across the fire. Their figures are blurred in the shot, only their profiles, illuminated by the firelight, their focused faces visible. The veteran explorer doesn't simply narrate silently; his movements are imbued with power: sometimes he uses his calloused right hand to gesture in the air to depict a treacherous terrain, sometimes he clenches his fist, as if reliving the tense moments of the past. Each gesture is precise and expressive. The primary light source for the entire scene comes from the blazing campfire in the center. The warm orange-red glow casts deep, shifting shadows on his face, separating him from the deep, cold coniferous forest nightscape and the star-studded sky behind him. This creates an atmosphere of both intense and captivating intimacy, as if the audience has become one of the listeners to this evening's conversation.", + "seed": 29845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29845_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": -0.01792535064741969, + "vlm_logprob": -0.01792535064741969, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.01792535064741969, + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028565604239702225 + ], + "opposite_logprob_tokens": [ + -3.6535656452178955 + ], + "logprob_reward": -0.028565604239702225, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028565604239702225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980854719877243, + 0.13073991239070892, + 0.10776247829198837, + 0.16023409366607666, + 0.04894239827990532, + 0.10350026935338974, + 0.06854046136140823, + 0.10016363114118576, + 0.04282481595873833, + 0.03920644149184227 + ], + "credit_map": [ + 0.039120230823755264, + 0.024606764316558838, + 0.015699660405516624, + 0.028303151950240135, + 0.02552465908229351, + 0.029317496344447136, + 0.02059563249349594, + 0.01491786539554596, + 0.01044300477951765, + 0.018305839970707893, + 0.010973143391311169, + 0.014066651463508606, + 0.023338906466960907, + 0.025879722088575363, + 0.01687372848391533, + 0.010858925059437752, + 0.006480577401816845, + 0.007665236946195364, + 0.00985414907336235, + 0.006384561304003, + 0.007658154238015413, + 0.029703043401241302, + 0.03525109216570854, + 0.004765658639371395, + 0.028294546529650688, + 0.01990663819015026, + 0.022069433704018593, + 0.018593065440654755, + 0.01947055757045746, + 0.02032136544585228, + 0.018456626683473587, + 0.013121851719915867, + 0.005346416495740414, + 0.0055131930857896805, + 0.0070208534598350525, + 0.0052689071744680405, + 0.0066805314272642136, + 0.006648826878517866, + 0.007038610056042671, + 0.005425056908279657, + 0.012565238401293755, + 0.011782035231590271, + 0.01463980134576559, + 0.01026915479451418, + 0.01793832890689373, + 0.013910834677517414, + 0.012347723357379436, + 0.010047157295048237, + 0.0076939878053963184, + 0.010233963839709759, + 0.008672475814819336, + 0.0061683389358222485, + 0.011087183840572834, + 0.010699490085244179, + 0.006420015823096037, + 0.0075650024227797985, + 0.009829417802393436, + 0.01769237592816353, + 0.013597779907286167, + 0.013241173699498177, + 0.012880906462669373, + 0.011056183837354183, + 0.01411303412169218, + 0.007752763107419014, + 0.0031787389889359474, + 0.007459379732608795, + 0.00603403989225626, + 0.0050230007618665695, + 0.005370863247662783, + 0.004410623572766781, + 0.006973940413445234, + 0.004374226555228233, + 0.003631529165431857, + 0.005499144550412893, + 0.005188353825360537, + 0.004739159252494574, + 0.005024314392358065, + 0.005575502756983042, + 0.004859099630266428, + 0.004689340479671955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.028565604239702225 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011094312183558941 + ], + "opposite_logprob_tokens": [ + -4.511094093322754 + ], + "logprob_reward": -0.011094312183558941, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011094312183558941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30708378553390503, + 0.14557236433029175, + 0.07322412729263306, + 0.09028400480747223, + 0.08203326910734177, + 0.05942515283823013, + 0.06770973652601242, + 0.049585502594709396, + 0.035819265991449356, + 0.08926279842853546 + ], + "credit_map": [ + 0.06877852976322174, + 0.038837991654872894, + 0.022854961454868317, + 0.03848939388990402, + 0.04390666261315346, + 0.04450559243559837, + 0.03053835593163967, + 0.019172312691807747, + 0.014790268614888191, + 0.017323767766356468, + 0.016706135123968124, + 0.01780928485095501, + 0.020210672169923782, + 0.0192081481218338, + 0.017545677721500397, + 0.021978408098220825, + 0.008450208231806755, + 0.008681496605277061, + 0.009262247011065483, + 0.010282769799232483, + 0.009967722930014133, + 0.008565370924770832, + 0.010959271341562271, + 0.007055043708533049, + 0.00780501076951623, + 0.0087570920586586, + 0.01642613857984543, + 0.007726889103651047, + 0.01159354392439127, + 0.010431737639009953, + 0.014495874755084515, + 0.013047718442976475, + 0.009892474859952927, + 0.007665221579372883, + 0.011894370429217815, + 0.010205622762441635, + 0.010010653175413609, + 0.009545940905809402, + 0.014198585413396358, + 0.008620399050414562, + 0.00960520375519991, + 0.008508195169270039, + 0.008450515568256378, + 0.005502511281520128, + 0.007185802794992924, + 0.008867934346199036, + 0.006083729211241007, + 0.005221258383244276, + 0.005818950477987528, + 0.009677737951278687, + 0.010613501071929932, + 0.006722493562847376, + 0.009959875606000423, + 0.008388055488467216, + 0.009255199693143368, + 0.007273920811712742, + 0.005503319203853607, + 0.010234668850898743, + 0.007006125058978796, + 0.006507714744657278, + 0.0057198056019842625, + 0.004951391369104385, + 0.00512296799570322, + 0.004539512563496828, + 0.0033438040409237146, + 0.006807516794651747, + 0.0059138089418411255, + 0.004330487456172705, + 0.004718868527561426, + 0.0036712063010782003, + 0.004410503897815943, + 0.0026230746880173683, + 0.007741307839751244, + 0.013079604133963585, + 0.01532311923801899, + 0.011833946220576763, + 0.009279310703277588, + 0.012267153710126877, + 0.011320464313030243, + 0.008417889475822449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011094312183558941 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550069019198418 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012550069019198418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550069019198418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17609119415283203, + 0.06654234975576401, + 0.08135082572698593, + 0.11001650243997574, + 0.09663768857717514, + 0.2737528085708618, + 0.0847383514046669, + 0.04108576476573944, + 0.027169762179255486, + 0.04261475428938866 + ], + "credit_map": [ + 0.021184444427490234, + 0.026054807007312775, + 0.015116536058485508, + 0.023401809856295586, + 0.03322525694966316, + 0.02539568580687046, + 0.01990327425301075, + 0.011809388175606728, + 0.006646996829658747, + 0.007684473879635334, + 0.008647174574434757, + 0.008547748439013958, + 0.011078627780079842, + 0.00939637329429388, + 0.007655776105821133, + 0.006885177921503782, + 0.009664228186011314, + 0.013589398935437202, + 0.008195338770747185, + 0.007525768131017685, + 0.01715860702097416, + 0.009941931813955307, + 0.009243891574442387, + 0.006031662225723267, + 0.022478323429822922, + 0.009392510168254375, + 0.018421052023768425, + 0.007773231249302626, + 0.01051242370158434, + 0.022308792918920517, + 0.011612346395850182, + 0.007517829071730375, + 0.016628142446279526, + 0.01120537519454956, + 0.013481567613780499, + 0.008174499496817589, + 0.010100885294377804, + 0.015019566752016544, + 0.011192390695214272, + 0.010835261084139347, + 0.02526475116610527, + 0.025614500045776367, + 0.036112040281295776, + 0.0387369729578495, + 0.033015333116054535, + 0.023703275248408318, + 0.035972606390714645, + 0.05533333122730255, + 0.00568738766014576, + 0.01138057466596365, + 0.010256893932819366, + 0.008954040706157684, + 0.011083167046308517, + 0.0082221245393157, + 0.011268761940300465, + 0.01788541115820408, + 0.003703298280015588, + 0.0054726311936974525, + 0.005084316246211529, + 0.0055854907259345055, + 0.004451938439160585, + 0.005445059854537249, + 0.005465631373226643, + 0.005877397954463959, + 0.002182604279369116, + 0.005076561123132706, + 0.004371054004877806, + 0.002854470629245043, + 0.003719668835401535, + 0.003420800669118762, + 0.0028891086112707853, + 0.0026554944925010204, + 0.0033128925133496523, + 0.006399593781679869, + 0.0057673160918056965, + 0.006390195805579424, + 0.005988179240375757, + 0.004970994777977467, + 0.005219461861997843, + 0.004566119518131018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.012550069019198418 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630132604390383 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630132604390383, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630132604390383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3120763301849365, + 0.2864835858345032, + 0.067458376288414, + 0.06937512010335922, + 0.09465128928422928, + 0.0730823203921318, + 0.02488335222005844, + 0.023076510056853294, + 0.019795171916484833, + 0.029117967933416367 + ], + "credit_map": [ + 0.032461587339639664, + 0.044116172939538956, + 0.04038580134510994, + 0.031119083985686302, + 0.04686383530497551, + 0.05085138976573944, + 0.03272700682282448, + 0.03355144336819649, + 0.025341637432575226, + 0.03185448795557022, + 0.038413818925619125, + 0.03835375979542732, + 0.042474519461393356, + 0.03341814503073692, + 0.029809696599841118, + 0.04681754484772682, + 0.0060201226733624935, + 0.0061215185560286045, + 0.009204147383570671, + 0.010832025669515133, + 0.009103277698159218, + 0.007388060912489891, + 0.010677087120711803, + 0.008112138137221336, + 0.006188482977449894, + 0.007335950154811144, + 0.010656655766069889, + 0.006796764675527811, + 0.012076915241777897, + 0.0065124607644975185, + 0.011379686184227467, + 0.008428208529949188, + 0.015216805972158909, + 0.008119795471429825, + 0.014831793494522572, + 0.00880307238548994, + 0.00909755565226078, + 0.017522737383842468, + 0.010277298279106617, + 0.010782231576740742, + 0.008530556224286556, + 0.007590252440422773, + 0.01161160133779049, + 0.0068145813420414925, + 0.009480656124651432, + 0.010707245208323002, + 0.00942170899361372, + 0.0089257238432765, + 0.002869651187211275, + 0.0036309564020484686, + 0.003631854197010398, + 0.0024369831662625074, + 0.003425868460908532, + 0.003303457982838154, + 0.0030438366811722517, + 0.0025407446082681417, + 0.0023181235883384943, + 0.0034071900881826878, + 0.0028435983695089817, + 0.0031107300892472267, + 0.003548309439793229, + 0.0029833263251930475, + 0.0029191591311246157, + 0.0019460741896182299, + 0.001940651796758175, + 0.002962066326290369, + 0.0029493693727999926, + 0.002128838561475277, + 0.0026507549919188023, + 0.002789779333397746, + 0.0023795210290700197, + 0.0019941912032663822, + 0.002594134770333767, + 0.004986860789358616, + 0.00409827521070838, + 0.003502113977447152, + 0.003834865754470229, + 0.0031654469203203917, + 0.004088960587978363, + 0.00284731131978333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007630132604390383 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978663519024849 + ], + "opposite_logprob_tokens": [ + -3.5297865867614746 + ], + "logprob_reward": -0.02978663519024849, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978663519024849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27431637048721313, + 0.11258218437433243, + 0.09090981632471085, + 0.09465979784727097, + 0.09390438348054886, + 0.13269735872745514, + 0.07024935632944107, + 0.05266842991113663, + 0.04081365093588829, + 0.037198614329099655 + ], + "credit_map": [ + 0.03819632530212402, + 0.036182574927806854, + 0.03337010741233826, + 0.024726154282689095, + 0.0519464835524559, + 0.041413091123104095, + 0.029731009155511856, + 0.018750615417957306, + 0.009834188967943192, + 0.01593121699988842, + 0.010690082795917988, + 0.01095146406441927, + 0.019312769174575806, + 0.021838512271642685, + 0.016060126945376396, + 0.007963817566633224, + 0.006837668363004923, + 0.006682421546429396, + 0.008215018548071384, + 0.010594316758215427, + 0.010803023353219032, + 0.020278647541999817, + 0.020937219262123108, + 0.006561505142599344, + 0.01275531854480505, + 0.011872133240103722, + 0.01381046324968338, + 0.01061814185231924, + 0.010616295970976353, + 0.014568369835615158, + 0.012925652787089348, + 0.007493422832340002, + 0.013697091490030289, + 0.008346255868673325, + 0.013183994218707085, + 0.011147607117891312, + 0.009788552299141884, + 0.010336702689528465, + 0.013985625468194485, + 0.013418554328382015, + 0.02389058843255043, + 0.011041641235351562, + 0.017611630260944366, + 0.01992223970592022, + 0.010842645540833473, + 0.014718788675963879, + 0.01642858050763607, + 0.018241234123706818, + 0.007208145689219236, + 0.010996523313224316, + 0.009380728006362915, + 0.006915245670825243, + 0.011948115192353725, + 0.010118053294718266, + 0.007510111667215824, + 0.0061724320985376835, + 0.005744547117501497, + 0.00922372005879879, + 0.008135281503200531, + 0.005400070454925299, + 0.006146952044218779, + 0.00825995672494173, + 0.005685092881321907, + 0.004072810057550669, + 0.0035620685666799545, + 0.007435887586325407, + 0.005479763727635145, + 0.005320445634424686, + 0.005823581945151091, + 0.0036475788801908493, + 0.005628699436783791, + 0.0039156218990683556, + 0.00333911064080894, + 0.0069097308441996574, + 0.004577444400638342, + 0.004299937281757593, + 0.005390237085521221, + 0.0035911703016608953, + 0.0056135267950594425, + 0.0034774590749293566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02978663519024849 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "documentary-style video, shot from a steady, handheld medium shot, held slightly upward, slowly and almost imperceptibly approaches its subject, focusing on an elderly explorer seated around a campfire. His face is weathered, deeply wrinkled like a dry riverbed, his bronzed skin gleaming oily in the dancing flames. An old scar runs from his left eyebrow to his temple. His graying hair is a little unkempt, and his thick beard quivers slightly with the rhythm of his speech. He wears a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest, frayed at the edges. He sits on a rough log, leaning slightly forward, his eyes sharp and focused, recounting past events to his companions across the fire. Their figures are blurred in the shot, only their profiles, illuminated by the firelight, their focused faces visible. The veteran explorer doesn't simply narrate silently; his movements are imbued with power: sometimes he uses his calloused right hand to gesture in the air to depict a treacherous terrain, sometimes he clenches his fist, as if reliving the tense moments of the past. Each gesture is precise and expressive. The primary light source for the entire scene comes from the blazing campfire in the center. The warm orange-red glow casts deep, shifting shadows on his face, separating him from the deep, cold coniferous forest nightscape and the star-studded sky behind him. This creates an atmosphere of both intense and captivating intimacy, as if the audience has become one of the listeners to this evening's conversation.", + "seed": 29846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_1/index_5510_seed_29846_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": -0.024786869529634714, + "vlm_logprob": -0.024786869529634714, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.024786869529634714, + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05458119511604309 + ], + "opposite_logprob_tokens": [ + -3.0545811653137207 + ], + "logprob_reward": -0.05458119511604309, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05458119511604309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21958443522453308, + 0.09505974501371384, + 0.07174286991357803, + 0.03813771530985832, + 0.13551858067512512, + 0.058008354157209396, + 0.05195070430636406, + 0.19963790476322174, + 0.0613907165825367, + 0.0689690038561821 + ], + "credit_map": [ + 0.020507384091615677, + 0.017404774203896523, + 0.03293349966406822, + 0.025401512160897255, + 0.03082279860973358, + 0.047486625611782074, + 0.025569729506969452, + 0.019458120688796043, + 0.008031954057514668, + 0.007523165550082922, + 0.015140462666749954, + 0.016440371051430702, + 0.009471321478486061, + 0.008796382695436478, + 0.018776342272758484, + 0.01087975688278675, + 0.006936944089829922, + 0.007012101821601391, + 0.010705229826271534, + 0.012161402963101864, + 0.007135773077607155, + 0.009917712770402431, + 0.010890867561101913, + 0.006982835941016674, + 0.0035781271290034056, + 0.004891333635896444, + 0.005301069002598524, + 0.004408801440149546, + 0.006434810347855091, + 0.004350784234702587, + 0.004251930862665176, + 0.004920859355479479, + 0.007850062102079391, + 0.00928493868559599, + 0.021457858383655548, + 0.017398538067936897, + 0.030198927968740463, + 0.01668696105480194, + 0.016453132033348083, + 0.016188159584999084, + 0.005530694965273142, + 0.006698078475892544, + 0.007382120005786419, + 0.011311476118862629, + 0.009265267290174961, + 0.006532068829983473, + 0.00606350414454937, + 0.005225144326686859, + 0.004362313076853752, + 0.00482881860807538, + 0.0075925616547465324, + 0.008735019713640213, + 0.0054832142777740955, + 0.009357800707221031, + 0.004370342940092087, + 0.007220634259283543, + 0.029329858720302582, + 0.024541199207305908, + 0.026987262070178986, + 0.04114328697323799, + 0.025484425947070122, + 0.017446763813495636, + 0.016885606572031975, + 0.017819497734308243, + 0.004468804225325584, + 0.005634089931845665, + 0.009920828975737095, + 0.011409484781324863, + 0.007054229732602835, + 0.005595157388597727, + 0.009509023278951645, + 0.007799101062119007, + 0.00498513737693429, + 0.005963170900940895, + 0.008926507085561752, + 0.011751302517950535, + 0.014036513864994049, + 0.005748808849602938, + 0.011958458460867405, + 0.005599105730652809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.05458119511604309 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052803573198616505 + ], + "opposite_logprob_tokens": [ + -5.255280494689941 + ], + "logprob_reward": -0.0052803573198616505, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052803573198616505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10307545214891434, + 0.0785420835018158, + 0.1270938515663147, + 0.06297322362661362, + 0.06121983006596565, + 0.045522864907979965, + 0.05709481239318848, + 0.3867061138153076, + 0.04757252335548401, + 0.03019927814602852 + ], + "credit_map": [ + 0.011600597761571407, + 0.010561715811491013, + 0.014482202008366585, + 0.014250939711928368, + 0.012248845770955086, + 0.01829005405306816, + 0.01283178012818098, + 0.00880932155996561, + 0.006195575464516878, + 0.01289340015500784, + 0.010554417036473751, + 0.012350930832326412, + 0.0075240228325128555, + 0.007502872962504625, + 0.011232154443860054, + 0.010288720950484276, + 0.00982600636780262, + 0.011603825725615025, + 0.015692073851823807, + 0.013916862197220325, + 0.016913700848817825, + 0.012487770989537239, + 0.028420062735676765, + 0.018233565613627434, + 0.004447312094271183, + 0.0045877681113779545, + 0.007815650664269924, + 0.010325985960662365, + 0.010230756364762783, + 0.008313885889947414, + 0.010576781816780567, + 0.006675083655864, + 0.004628258757293224, + 0.0044905380345880985, + 0.010606222786009312, + 0.0168819148093462, + 0.006288496311753988, + 0.0055656228214502335, + 0.007048339117318392, + 0.0057104420848190784, + 0.009418560191988945, + 0.008582433685660362, + 0.006929459515959024, + 0.004611032083630562, + 0.003482622792944312, + 0.00368303619325161, + 0.002860589884221554, + 0.005955133121460676, + 0.004739497322589159, + 0.005625370889902115, + 0.007944181561470032, + 0.009796143509447575, + 0.00611048424616456, + 0.009681690484285355, + 0.0057435184717178345, + 0.007453925907611847, + 0.03957078605890274, + 0.04352124035358429, + 0.052626386284828186, + 0.08398500829935074, + 0.05593128502368927, + 0.03624897450208664, + 0.03553655371069908, + 0.03928586468100548, + 0.009206393733620644, + 0.006293858867138624, + 0.006798864342272282, + 0.005694730207324028, + 0.004738153889775276, + 0.005163709167391062, + 0.004897632170468569, + 0.004779185634106398, + 0.0029281466268002987, + 0.003678901121020317, + 0.003901578951627016, + 0.003811294911429286, + 0.004529168363660574, + 0.004707920830696821, + 0.003541863290593028, + 0.003100405912846327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0052803573198616505 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820743829011917 + ], + "opposite_logprob_tokens": [ + -4.018207550048828 + ], + "logprob_reward": -0.01820743829011917, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820743829011917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.123881496489048, + 0.07461833953857422, + 0.0727330669760704, + 0.025252893567085266, + 0.10174785554409027, + 0.10901731997728348, + 0.08441236615180969, + 0.3416680097579956, + 0.045845091342926025, + 0.02082354947924614 + ], + "credit_map": [ + 0.011767851188778877, + 0.016101829707622528, + 0.0187687985599041, + 0.018357498571276665, + 0.016369661316275597, + 0.014074469916522503, + 0.016552027314901352, + 0.011889367364346981, + 0.005844659171998501, + 0.008513977751135826, + 0.010277151130139828, + 0.01029436755925417, + 0.010939677245914936, + 0.006777952425181866, + 0.01062326692044735, + 0.011347286403179169, + 0.005309777334332466, + 0.006415162701159716, + 0.010274253785610199, + 0.008992895483970642, + 0.00775501923635602, + 0.0088370805606246, + 0.017822548747062683, + 0.007326332852244377, + 0.0022252683993428946, + 0.0038471028674393892, + 0.0034682955592870712, + 0.0034535250160843134, + 0.002359678503125906, + 0.0028700788971036673, + 0.00398296257480979, + 0.0030459819827228785, + 0.009538939222693443, + 0.006817801855504513, + 0.01711246930062771, + 0.022751087322831154, + 0.014466037042438984, + 0.010380257852375507, + 0.01098716352134943, + 0.00969410315155983, + 0.008905082009732723, + 0.012098033912479877, + 0.015918441116809845, + 0.02911296673119068, + 0.010751307010650635, + 0.01316260639578104, + 0.008975688368082047, + 0.010093193501234055, + 0.006241369526833296, + 0.007339369971305132, + 0.012213800102472305, + 0.020393870770931244, + 0.008451775647699833, + 0.013194513507187366, + 0.006717532407492399, + 0.009860132820904255, + 0.04182618111371994, + 0.03664001077413559, + 0.045552387833595276, + 0.06981150060892105, + 0.0496758297085762, + 0.03497833386063576, + 0.03031228482723236, + 0.03287147358059883, + 0.004792207386344671, + 0.006752108223736286, + 0.007183223031461239, + 0.007525949738919735, + 0.00555496895685792, + 0.004456335678696632, + 0.004069535061717033, + 0.005510765593498945, + 0.0016128550050780177, + 0.0016695247031748295, + 0.0031165925320237875, + 0.00360872526653111, + 0.0036166037898510695, + 0.0031307118479162455, + 0.001871987828053534, + 0.0021965496707707644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01820743829011917 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072357073426247 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072357073426247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072357073426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24378234148025513, + 0.15432782471179962, + 0.10127802938222885, + 0.051480598747730255, + 0.11085361242294312, + 0.07279334217309952, + 0.06547920405864716, + 0.11481542885303497, + 0.06570287048816681, + 0.01948675885796547 + ], + "credit_map": [ + 0.032967470586299896, + 0.0238333847373724, + 0.03310154378414154, + 0.030366873368620872, + 0.031950049102306366, + 0.036606624722480774, + 0.02815237268805504, + 0.026804011315107346, + 0.009758858941495419, + 0.011702381074428558, + 0.022257518023252487, + 0.01947844959795475, + 0.023006906732916832, + 0.017316261306405067, + 0.028049994260072708, + 0.022757433354854584, + 0.009654290974140167, + 0.008843210525810719, + 0.013813463971018791, + 0.018123803660273552, + 0.011453074403107166, + 0.014648184180259705, + 0.015398208983242512, + 0.009343783371150494, + 0.0026405879762023687, + 0.0040252916514873505, + 0.007586922030895948, + 0.012019844725728035, + 0.005878489930182695, + 0.007448537275195122, + 0.006798258982598782, + 0.0050826603546738625, + 0.02240505814552307, + 0.013083978556096554, + 0.014864050783216953, + 0.019620250910520554, + 0.008554023690521717, + 0.007907884195446968, + 0.009406647644937038, + 0.015011713840067387, + 0.010534923523664474, + 0.010230652987957, + 0.009546211920678616, + 0.01350482553243637, + 0.00985038559883833, + 0.006061432417482138, + 0.00666527496650815, + 0.0063996328972280025, + 0.005474186036735773, + 0.006771602667868137, + 0.008617492392659187, + 0.009716523811221123, + 0.00876864418387413, + 0.009231680072844028, + 0.0077600060030817986, + 0.009139062836766243, + 0.008131301030516624, + 0.011485965922474861, + 0.016726380214095116, + 0.02733507938683033, + 0.01494365744292736, + 0.012647467665374279, + 0.011960601434111595, + 0.011584969237446785, + 0.007535878568887711, + 0.009999704547226429, + 0.008951048366725445, + 0.009560192935168743, + 0.0073683965019881725, + 0.007642536424100399, + 0.008043157868087292, + 0.0066019524820148945, + 0.0020494319032877684, + 0.0014823894016444683, + 0.0026252283714711666, + 0.002644044579938054, + 0.0027333067264407873, + 0.002708819694817066, + 0.003093813080340624, + 0.002149723470211029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016072357073426247 + }, + { + "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 explorer with weathered, deeply wrinkled skin and a scar running from his left eyebrow to his temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a thick, washed-out flannel plaid shirt over a multi-pocketed dark brown canvas vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the explorer use his right hand to make precise gestures, such as depicting terrain or clenching his fist, while speaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit primarily by a blazing campfire in the center, casting warm orange-red glow and deep shifting shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld, medium, held slightly upward, and slowly approaching the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979299984872341 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.02979299984872341, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979299984872341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17930066585540771, + 0.10032805800437927, + 0.09690394997596741, + 0.034453388303518295, + 0.09739071875810623, + 0.12269826978445053, + 0.09300809353590012, + 0.13890792429447174, + 0.098101906478405, + 0.0389070026576519 + ], + "credit_map": [ + 0.014668024145066738, + 0.01723589561879635, + 0.02592509426176548, + 0.03744427114725113, + 0.019477859139442444, + 0.028471006080508232, + 0.017269061878323555, + 0.018809469416737556, + 0.009462270885705948, + 0.011547810398042202, + 0.015419100411236286, + 0.018481537699699402, + 0.012787336483597755, + 0.009097978472709656, + 0.012343374080955982, + 0.01118865329772234, + 0.007712446618825197, + 0.007920199073851109, + 0.014268143102526665, + 0.014390278607606888, + 0.011666499078273773, + 0.010274238884449005, + 0.021461868658661842, + 0.009210282005369663, + 0.002806740812957287, + 0.0048937490209937096, + 0.00546829029917717, + 0.004829895216971636, + 0.0056058140471577644, + 0.0034286268055438995, + 0.004765388555824757, + 0.002654886106029153, + 0.00770958885550499, + 0.00827307254076004, + 0.013795116916298866, + 0.016467375680804253, + 0.015912435948848724, + 0.009924291633069515, + 0.014936672523617744, + 0.010372176766395569, + 0.02303503267467022, + 0.021373657509684563, + 0.018841098994016647, + 0.017284424975514412, + 0.01071113720536232, + 0.007628452964127064, + 0.009671718813478947, + 0.01415274664759636, + 0.008165618404746056, + 0.01160054374486208, + 0.01438824087381363, + 0.01877741515636444, + 0.009139465168118477, + 0.00984885636717081, + 0.009602866135537624, + 0.011485086753964424, + 0.00928979367017746, + 0.015041389502584934, + 0.019348207861185074, + 0.024515023455023766, + 0.022849196568131447, + 0.019506286829710007, + 0.013326581567525864, + 0.015031438320875168, + 0.009455091319978237, + 0.014360816217958927, + 0.014293002896010876, + 0.015255811624228954, + 0.011499358341097832, + 0.010196025483310223, + 0.011662222445011139, + 0.01137958187609911, + 0.004317684564739466, + 0.004743867553770542, + 0.004819538444280624, + 0.005037515424191952, + 0.005680563859641552, + 0.004835912492126226, + 0.005313604138791561, + 0.004158316645771265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.02979299984872341 + } + ], + "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/0000017/group_10/index_tensor([7129, 7129])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_tensor([7129, 7129])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..61547ab11d336e324ef4fac39558729796a13d74 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_tensor([7129, 7129])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly creative first-person point-of-view shot, shot from inside the gaping mouth of a Predator warhound, its jagged, rounded teeth providing a natural frame for the frame, while its pink gums and the occasional wet tongue flickering across the camera add chaotic detail. The camera focuses on an equally goofy centaur, grappling with the warhound in a comical struggle. The centaur, sporting an exaggerated, comical expression of both panic and excitement, struggles to free itself by pushing against the warhound's face with both arms. In a childishly animated aesthetic, both characters are rendered with bold black outlines and vibrant, flat colors, their movements exaggerated and elastic. The camera shakes violently and chaotically as the warhound's head sways and the two strugglers wrestle, perfectly capturing the dynamics of the melee. The backdrop is a minimalist, brightly lit landscape of grass and blue sky, with even lighting emphasizing the scene's stylized quality. Ultimately, this scuffle, recorded from a unique perspective, creates an absurd, chaotic, and childlike atmosphere through its unstable images and the characters' clumsy interactions.", + "seed": 6655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6655_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": -0.06965017057955265, + "vlm_logprob": -0.06965017057955265, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.06965017057955265, + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.015074852854013443 + ], + "opposite_logprob_tokens": [ + -4.640074729919434 + ], + "logprob_reward": -0.015074852854013443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015074852854013443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1194484606385231, + 0.04206184670329094, + 0.14615333080291748, + 0.1795073300600052, + 0.05838845670223236, + 0.061299558728933334, + 0.14668548107147217, + 0.03575202450156212, + 0.0681075006723404, + 0.14259599149227142 + ], + "credit_map": [ + 0.009517590515315533, + 0.017406858503818512, + 0.012414297088980675, + 0.01879449561238289, + 0.012340078130364418, + 0.01099400408565998, + 0.028934713453054428, + 0.009046425111591816, + 0.0037763139698654413, + 0.005616873037070036, + 0.0058140745386481285, + 0.007385631091892719, + 0.006415660493075848, + 0.0033731081057339907, + 0.004687707405537367, + 0.00499247619882226, + 0.008431412279605865, + 0.007891076616942883, + 0.018376104533672333, + 0.007730900309979916, + 0.0059666866436600685, + 0.007020274177193642, + 0.040437959134578705, + 0.05029892548918724, + 0.009036248549818993, + 0.004744693171232939, + 0.06461907178163528, + 0.010157481767237186, + 0.006551510654389858, + 0.06610587984323502, + 0.008413849398493767, + 0.009878592565655708, + 0.010462800972163677, + 0.006773046683520079, + 0.012075399979948997, + 0.008962784893810749, + 0.004367891699075699, + 0.005593622103333473, + 0.004805936943739653, + 0.00534697575494647, + 0.006030477117747068, + 0.005350957158952951, + 0.006352052092552185, + 0.00810767151415348, + 0.011374291032552719, + 0.009249514900147915, + 0.00879778154194355, + 0.006036817561835051, + 0.011313407681882381, + 0.009371474385261536, + 0.012837669812142849, + 0.04834726080298424, + 0.010081200860440731, + 0.04049667716026306, + 0.00886278785765171, + 0.00537500623613596, + 0.0037173477467149496, + 0.004346201661974192, + 0.00609363941475749, + 0.003854120150208473, + 0.004389442037791014, + 0.0034652245230972767, + 0.005949073936790228, + 0.003936978057026863, + 0.00728578818961978, + 0.008914457634091377, + 0.009660578332841396, + 0.010101952590048313, + 0.00774970231577754, + 0.006848969496786594, + 0.007410344202071428, + 0.01013569813221693, + 0.009619718417525291, + 0.014041494578123093, + 0.017313020303845406, + 0.03618505224585533, + 0.011508895084261894, + 0.0307576023042202, + 0.012702640146017075, + 0.01046757958829403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.015074852854013443 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.02634584717452526 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634584717452526, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634584717452526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13990335166454315, + 0.037692852318286896, + 0.2814045548439026, + 0.15604542195796967, + 0.027404336258769035, + 0.05108720436692238, + 0.1591341346502304, + 0.028680583462119102, + 0.05468931049108505, + 0.0639583095908165 + ], + "credit_map": [ + 0.0072158752009272575, + 0.010224215686321259, + 0.013858696445822716, + 0.013794885948300362, + 0.012658875435590744, + 0.018354400992393494, + 0.04828936234116554, + 0.015507030300796032, + 0.0035340492613613605, + 0.005149465054273605, + 0.004469321109354496, + 0.004704711027443409, + 0.00478557450696826, + 0.003843548707664013, + 0.007127693388611078, + 0.004078488796949387, + 0.016961945220828056, + 0.011058671399950981, + 0.0399322435259819, + 0.014304225333034992, + 0.010262973606586456, + 0.017844362184405327, + 0.07298405468463898, + 0.09805605560541153, + 0.008950570598244667, + 0.0037184355314821005, + 0.05410921573638916, + 0.009115176275372505, + 0.006002272479236126, + 0.05323910340666771, + 0.009357639588415623, + 0.011553000658750534, + 0.004655162803828716, + 0.004332754295319319, + 0.00406281603500247, + 0.0025641480460762978, + 0.002537763211876154, + 0.003995431121438742, + 0.0028810224030166864, + 0.002375238109380007, + 0.0051062447018921375, + 0.003851786255836487, + 0.004452126566320658, + 0.007108103018254042, + 0.008623801171779633, + 0.008146832697093487, + 0.007320021279156208, + 0.006478280760347843, + 0.011960670351982117, + 0.010697728022933006, + 0.011228512972593307, + 0.05270998552441597, + 0.011790660209953785, + 0.04780134558677673, + 0.00733394268900156, + 0.005611267872154713, + 0.0025184263940900564, + 0.0021820557303726673, + 0.005307838320732117, + 0.0031183345708996058, + 0.004367349203675985, + 0.0028770160861313343, + 0.004424880258738995, + 0.003884680336341262, + 0.005491519346833229, + 0.00542091578245163, + 0.004280691966414452, + 0.004186443518847227, + 0.012310213409364223, + 0.009221339598298073, + 0.006613101344555616, + 0.007165079470723867, + 0.004232337232679129, + 0.005754996091127396, + 0.007441412657499313, + 0.015317469835281372, + 0.005916595924645662, + 0.01480026450008154, + 0.006250672973692417, + 0.004244557581841946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.02634584717452526 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.02632889524102211 + ], + "opposite_logprob_tokens": [ + -3.6513288021087646 + ], + "logprob_reward": -0.02632889524102211, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632889524102211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22259850800037384, + 0.04862936586141586, + 0.11706564575433731, + 0.4439932703971863, + 0.024086346849799156, + 0.032202064990997314, + 0.04288027063012123, + 0.01859571784734726, + 0.017605431377887726, + 0.032343339174985886 + ], + "credit_map": [ + 0.011173229664564133, + 0.01639673113822937, + 0.020261088386178017, + 0.02004207856953144, + 0.019141407683491707, + 0.03046003170311451, + 0.08559556305408478, + 0.01952839083969593, + 0.0039932155050337315, + 0.009239977225661278, + 0.0053544114343822, + 0.006788919679820538, + 0.005271201021969318, + 0.0030966559424996376, + 0.010560635477304459, + 0.004324354697018862, + 0.012362631969153881, + 0.009831658564507961, + 0.005938713438808918, + 0.012137478217482567, + 0.006799545604735613, + 0.00812047440558672, + 0.04378676787018776, + 0.01808837801218033, + 0.02159530483186245, + 0.010425282642245293, + 0.15673598647117615, + 0.02639833651483059, + 0.018380703404545784, + 0.15965381264686584, + 0.02027839608490467, + 0.03052545338869095, + 0.002243679016828537, + 0.0027573187835514545, + 0.003657870925962925, + 0.004088096786290407, + 0.002858849475160241, + 0.002989137312397361, + 0.003154174890369177, + 0.002337220823392272, + 0.003652912098914385, + 0.002586033660918474, + 0.00384219316765666, + 0.003956136759370565, + 0.0056487624533474445, + 0.004781683906912804, + 0.004856516141444445, + 0.002877830760553479, + 0.0038746134378015995, + 0.005441324785351753, + 0.0037808846682310104, + 0.009067486971616745, + 0.005619534291327, + 0.010526733472943306, + 0.002674999414011836, + 0.0018946921918541193, + 0.0014347507385537028, + 0.002760190749540925, + 0.003290640190243721, + 0.002932377392426133, + 0.0018098469590768218, + 0.0016871544066816568, + 0.0019934365991503, + 0.002687321975827217, + 0.0016374036204069853, + 0.0015373745700344443, + 0.002440118230879307, + 0.0018924648175016046, + 0.003383693052455783, + 0.002727989573031664, + 0.0020204565953463316, + 0.0019659327808767557, + 0.003100466215983033, + 0.0036230157129466534, + 0.003900783136487007, + 0.007034737151116133, + 0.002437238348647952, + 0.005486654117703438, + 0.0035784796345978975, + 0.003181968815624714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.02632889524102211 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.0549444817006588 + ], + "opposite_logprob_tokens": [ + -2.9299445152282715 + ], + "logprob_reward": -0.0549444817006588, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549444817006588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12972307205200195, + 0.04246656596660614, + 0.36649081110954285, + 0.15111388266086578, + 0.065196193754673, + 0.04296677187085152, + 0.051171157509088516, + 0.0793394073843956, + 0.021173624321818352, + 0.05035850778222084 + ], + "credit_map": [ + 0.009793211705982685, + 0.013738220557570457, + 0.01051692757755518, + 0.013137497007846832, + 0.012781348079442978, + 0.019023269414901733, + 0.04004635289311409, + 0.010686246678233147, + 0.003850162960588932, + 0.005669968202710152, + 0.00429897103458643, + 0.006192455533891916, + 0.00657686498016119, + 0.005309640895575285, + 0.0047806971706449986, + 0.005787805654108524, + 0.021344749256968498, + 0.009751429781317711, + 0.05673068389296532, + 0.01601516827940941, + 0.010436778888106346, + 0.012202996760606766, + 0.0918060764670372, + 0.14820292592048645, + 0.006842270493507385, + 0.0035945852287113667, + 0.05391685664653778, + 0.00856013223528862, + 0.005056232213973999, + 0.05487973615527153, + 0.010078778490424156, + 0.008185288868844509, + 0.010484009981155396, + 0.00930256862193346, + 0.013041451573371887, + 0.008270814083516598, + 0.004980354104191065, + 0.008180209435522556, + 0.005926206707954407, + 0.005010576918721199, + 0.004704040940850973, + 0.003936107270419598, + 0.004619952291250229, + 0.006114044226706028, + 0.00712828803807497, + 0.006243797484785318, + 0.006204682402312756, + 0.004015863873064518, + 0.0036655552685260773, + 0.004519252106547356, + 0.00506123760715127, + 0.0131841991096735, + 0.006824774667620659, + 0.009664448909461498, + 0.005239800550043583, + 0.00301189417950809, + 0.008318104781210423, + 0.01165720447897911, + 0.015802793204784393, + 0.009150431491434574, + 0.007630011532455683, + 0.007943863049149513, + 0.011381382122635841, + 0.007455614861100912, + 0.0023910729214549065, + 0.001684731338173151, + 0.003221415914595127, + 0.0018793706549331546, + 0.0038358676247298717, + 0.003480251645669341, + 0.002233291743323207, + 0.0024476221296936274, + 0.0037492304109036922, + 0.005443132482469082, + 0.007119411136955023, + 0.01278550922870636, + 0.0037209189031273127, + 0.009448016062378883, + 0.004831009544432163, + 0.003261280944570899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.0549444817006588 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.22555677592754364 + ], + "opposite_logprob_tokens": [ + -1.60055673122406 + ], + "logprob_reward": -0.22555677592754364, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555677592754364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18432733416557312, + 0.021826641634106636, + 0.484346866607666, + 0.18659472465515137, + 0.010347679257392883, + 0.01878899522125721, + 0.043602775782346725, + 0.011427300982177258, + 0.009232626296579838, + 0.029505038633942604 + ], + "credit_map": [ + 0.007188646122813225, + 0.009938950650393963, + 0.015424893237650394, + 0.018189886584877968, + 0.010396449826657772, + 0.033549826592206955, + 0.0722455084323883, + 0.017393171787261963, + 0.0018400533590465784, + 0.0029218257404863834, + 0.0030089342035353184, + 0.0037506178487092257, + 0.004145665094256401, + 0.001236104522831738, + 0.0025397930294275284, + 0.002383647719398141, + 0.027832109481096268, + 0.015897804871201515, + 0.06586938351392746, + 0.023687176406383514, + 0.016281310468912125, + 0.022329939529299736, + 0.13230960071086884, + 0.18013955652713776, + 0.010800518095493317, + 0.008423665538430214, + 0.05532027408480644, + 0.013213543221354485, + 0.01008154358714819, + 0.057606227695941925, + 0.013726143166422844, + 0.01742282137274742, + 0.0014280503382906318, + 0.0010387003421783447, + 0.0018744966946542263, + 0.0012453868985176086, + 0.0013145998818799853, + 0.0014436467317864299, + 0.0009865284664556384, + 0.0010162701364606619, + 0.0018635246669873595, + 0.0015538634033873677, + 0.0019142987439408898, + 0.002678101882338524, + 0.003540776902809739, + 0.00305444048717618, + 0.0023910189047455788, + 0.0017929708119481802, + 0.002679685829207301, + 0.0025904064532369375, + 0.0037712547928094864, + 0.010536915622651577, + 0.005833637900650501, + 0.009735860861837864, + 0.005678740330040455, + 0.0027762760873883963, + 0.0011273740092292428, + 0.0014585305470973253, + 0.0015899613499641418, + 0.0013625714927911758, + 0.0016546394908800721, + 0.0009265054250136018, + 0.001663990318775177, + 0.0016437276499345899, + 0.000840619730297476, + 0.0009605895029380918, + 0.0008826383273117244, + 0.0010082426015287638, + 0.002095583826303482, + 0.0015067162457853556, + 0.0007876110030338168, + 0.0011506250593811274, + 0.0028675622306764126, + 0.0022710980847477913, + 0.0029451583977788687, + 0.007931165397167206, + 0.002631769049912691, + 0.005566547159105539, + 0.002914423355832696, + 0.002377315191552043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.22555677592754364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly creative first-person point-of-view shot, shot from inside the gaping mouth of a Predator warhound, its jagged, rounded teeth providing a natural frame for the frame, while its pink gums and the occasional wet tongue flickering across the camera add chaotic detail. The camera focuses on an equally goofy centaur, grappling with the warhound in a comical struggle. The centaur, sporting an exaggerated, comical expression of both panic and excitement, struggles to free itself by pushing against the warhound's face with both arms. In a childishly animated aesthetic, both characters are rendered with bold black outlines and vibrant, flat colors, their movements exaggerated and elastic. The camera shakes violently and chaotically as the warhound's head sways and the two strugglers wrestle, perfectly capturing the dynamics of the melee. The backdrop is a minimalist, brightly lit landscape of grass and blue sky, with even lighting emphasizing the scene's stylized quality. Ultimately, this scuffle, recorded from a unique perspective, creates an absurd, chaotic, and childlike atmosphere through its unstable images and the characters' clumsy interactions.", + "seed": 6656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6656_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": -0.0326448243111372, + "vlm_logprob": -0.0326448243111372, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.0326448243111372, + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.035762447863817215 + ], + "opposite_logprob_tokens": [ + -3.5357625484466553 + ], + "logprob_reward": -0.035762447863817215, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035762447863817215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2045462727546692, + 0.10886114835739136, + 0.09639335423707962, + 0.09413284063339233, + 0.07974844425916672, + 0.057069551199674606, + 0.07875107228755951, + 0.04489695280790329, + 0.07147715240716934, + 0.1641233265399933 + ], + "credit_map": [ + 0.023892713710665703, + 0.018643515184521675, + 0.031503237783908844, + 0.02978592738509178, + 0.02449086681008339, + 0.020882906392216682, + 0.020617587491869926, + 0.034729503095149994, + 0.009542782790958881, + 0.028312934562563896, + 0.006424053106456995, + 0.020178547129034996, + 0.0119111193343997, + 0.011473948135972023, + 0.00796855054795742, + 0.013049200177192688, + 0.010502893477678299, + 0.006236524786800146, + 0.006239602342247963, + 0.010499967262148857, + 0.009134314022958279, + 0.015432441607117653, + 0.006242737174034119, + 0.032104868441820145, + 0.009869767352938652, + 0.017199328169226646, + 0.007360368501394987, + 0.018775442615151405, + 0.01136390957981348, + 0.010762079618871212, + 0.010910003446042538, + 0.007891938090324402, + 0.006947681773453951, + 0.00490392092615366, + 0.008441552519798279, + 0.016180817037820816, + 0.004057878628373146, + 0.009535737335681915, + 0.00802159495651722, + 0.021659253165125847, + 0.005747658666223288, + 0.009986955672502518, + 0.006196891888976097, + 0.008330648764967918, + 0.0054384018294513226, + 0.005140279419720173, + 0.008022375404834747, + 0.008206333965063095, + 0.005944617558270693, + 0.006259443704038858, + 0.006203108467161655, + 0.007693619001656771, + 0.009719260968267918, + 0.022118715569376945, + 0.01382253598421812, + 0.006989757064729929, + 0.006706428714096546, + 0.004474836401641369, + 0.0068647600710392, + 0.006573187652975321, + 0.0058035231195390224, + 0.0034502120688557625, + 0.005987039767205715, + 0.005036958958953619, + 0.006506500765681267, + 0.0063257706351578236, + 0.00572446221485734, + 0.008177237585186958, + 0.0068210032768547535, + 0.019762469455599785, + 0.010584844276309013, + 0.007574861403554678, + 0.01687365211546421, + 0.022339986637234688, + 0.016074422746896744, + 0.022492462769150734, + 0.019604794681072235, + 0.020950695499777794, + 0.025373106822371483, + 0.02041417732834816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.035762447863817215 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.029802486300468445 + ], + "opposite_logprob_tokens": [ + -3.5298025608062744 + ], + "logprob_reward": -0.029802486300468445, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802486300468445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1093214750289917, + 0.052424684166908264, + 0.11208038777112961, + 0.1619999259710312, + 0.12258078157901764, + 0.06938651949167252, + 0.03576859459280968, + 0.04915986955165863, + 0.10778060555458069, + 0.1794971078634262 + ], + "credit_map": [ + 0.016235629096627235, + 0.012686173431575298, + 0.018369659781455994, + 0.013803957961499691, + 0.013059960678219795, + 0.00912628322839737, + 0.01128145307302475, + 0.014758361503481865, + 0.00514989485964179, + 0.007157052401453257, + 0.004065284971147776, + 0.010145009495317936, + 0.007257661782205105, + 0.006065518595278263, + 0.004913628101348877, + 0.007670636288821697, + 0.012432454153895378, + 0.008475661277770996, + 0.009766998700797558, + 0.01360042579472065, + 0.012702247127890587, + 0.014433858916163445, + 0.009276673197746277, + 0.03139206022024155, + 0.015352473594248295, + 0.02327732928097248, + 0.0113041577860713, + 0.020204460248351097, + 0.022404268383979797, + 0.02926972135901451, + 0.02067551389336586, + 0.01951201818883419, + 0.006362879648804665, + 0.005873274058103561, + 0.006531702354550362, + 0.010252879001200199, + 0.004975734278559685, + 0.04402872174978256, + 0.008957965299487114, + 0.03559763357043266, + 0.007133578881621361, + 0.011162946932017803, + 0.007464640773832798, + 0.010467876680195332, + 0.007039977237582207, + 0.005673081614077091, + 0.009892430156469345, + 0.01055198349058628, + 0.0037339022383093834, + 0.0031314862426370382, + 0.0030147850047796965, + 0.003959817346185446, + 0.003980523440986872, + 0.00862428080290556, + 0.006139056757092476, + 0.0031847450882196426, + 0.0057064443826675415, + 0.005206787493079901, + 0.004608965013176203, + 0.005790413357317448, + 0.006619160529226065, + 0.004607678856700659, + 0.01017054170370102, + 0.0064498791471123695, + 0.010114242322742939, + 0.008004922419786453, + 0.012375994585454464, + 0.012499270029366016, + 0.010335277765989304, + 0.027144983410835266, + 0.014907799661159515, + 0.012398114427924156, + 0.020923543721437454, + 0.022518981248140335, + 0.017659427598118782, + 0.021865323185920715, + 0.01938503421843052, + 0.023017974570393562, + 0.02646893821656704, + 0.027657892554998398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.029802486300468445 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.02978455275297165 + ], + "opposite_logprob_tokens": [ + -3.5297844409942627 + ], + "logprob_reward": -0.02978455275297165, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978455275297165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15915587544441223, + 0.06704962998628616, + 0.1133783608675003, + 0.06675286591053009, + 0.2823168635368347, + 0.07080282270908356, + 0.061989106237888336, + 0.06195450201630592, + 0.06789683550596237, + 0.048703134059906006 + ], + "credit_map": [ + 0.016198517754673958, + 0.015681231394410133, + 0.017190828919410706, + 0.019660213962197304, + 0.012697096914052963, + 0.021649518981575966, + 0.01554692629724741, + 0.040531545877456665, + 0.005763357039541006, + 0.010018037632107735, + 0.005334462504833937, + 0.01788547821342945, + 0.005570041481405497, + 0.006897089071571827, + 0.006480904761701822, + 0.009100258350372314, + 0.010768501088023186, + 0.011969774961471558, + 0.009008503518998623, + 0.014278040267527103, + 0.01333557441830635, + 0.015769513323903084, + 0.006499672308564186, + 0.03174877539277077, + 0.006931716110557318, + 0.012081652879714966, + 0.0056211031042039394, + 0.009857957251369953, + 0.006668472196906805, + 0.006262834183871746, + 0.011689737439155579, + 0.00763939693570137, + 0.012471902184188366, + 0.013256382197141647, + 0.01154672633856535, + 0.0194670669734478, + 0.010324500501155853, + 0.10235172510147095, + 0.017158465459942818, + 0.09574007987976074, + 0.009934230707585812, + 0.00990570429712534, + 0.007075342815369368, + 0.013174314983189106, + 0.0072267609648406506, + 0.005944178905338049, + 0.008578293956816196, + 0.008963996544480324, + 0.005584227852523327, + 0.004764072131365538, + 0.006611856631934643, + 0.005852693226188421, + 0.00825770664960146, + 0.014244056306779385, + 0.010505367070436478, + 0.006169123109430075, + 0.008335378021001816, + 0.005586062557995319, + 0.007461773231625557, + 0.006396269425749779, + 0.00928217638283968, + 0.011768510565161705, + 0.007420708891004324, + 0.005703622940927744, + 0.007447863928973675, + 0.006566158961504698, + 0.010396366938948631, + 0.008454899303615093, + 0.0066076419316232204, + 0.011789511889219284, + 0.007282659411430359, + 0.009351733140647411, + 0.0044761416502296925, + 0.007016569841653109, + 0.004194481763988733, + 0.007121390663087368, + 0.005027128849178553, + 0.005421340465545654, + 0.006109051872044802, + 0.009337027557194233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.02978455275297165 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.038088809698820114 + ], + "opposite_logprob_tokens": [ + -3.288088798522949 + ], + "logprob_reward": -0.038088809698820114, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088809698820114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09843776375055313, + 0.05381843075156212, + 0.12068543583154678, + 0.08554232120513916, + 0.14831051230430603, + 0.19619688391685486, + 0.064546599984169, + 0.1430758386850357, + 0.026877835392951965, + 0.06250843405723572 + ], + "credit_map": [ + 0.011562705039978027, + 0.0096674095839262, + 0.01832561567425728, + 0.015089216642081738, + 0.00803653709590435, + 0.00883317831903696, + 0.010720825754106045, + 0.016202278435230255, + 0.004466434475034475, + 0.009111292660236359, + 0.005245268810540438, + 0.014656633138656616, + 0.004072696901857853, + 0.0054079038091003895, + 0.004493955988436937, + 0.006364244036376476, + 0.012462210841476917, + 0.0099647156894207, + 0.008705607615411282, + 0.015020519495010376, + 0.012923229485750198, + 0.0157304834574461, + 0.009776495397090912, + 0.03610217198729515, + 0.0076416656374931335, + 0.0072723315097391605, + 0.0056503890082240105, + 0.01706700585782528, + 0.009640837088227272, + 0.00916405487805605, + 0.017166776582598686, + 0.011939259245991707, + 0.01125780027359724, + 0.007343714591115713, + 0.009729092009365559, + 0.017427682876586914, + 0.005595777183771133, + 0.03166786581277847, + 0.012682007625699043, + 0.05260657146573067, + 0.02589298225939274, + 0.028162695467472076, + 0.021642878651618958, + 0.03821174427866936, + 0.022007765248417854, + 0.018394162878394127, + 0.022763654589653015, + 0.019120994955301285, + 0.009162134490907192, + 0.004833414684981108, + 0.013098081573843956, + 0.008956905454397202, + 0.004941388498991728, + 0.006245225202292204, + 0.009936015121638775, + 0.007373435888439417, + 0.022402167320251465, + 0.02107975259423256, + 0.015722982585430145, + 0.023126916959881783, + 0.0153966648504138, + 0.016391901299357414, + 0.01149599626660347, + 0.01745946891605854, + 0.0030358461663126945, + 0.0026989425532519817, + 0.003140727523714304, + 0.0028451024554669857, + 0.0022247631568461657, + 0.004655236843973398, + 0.004832539241760969, + 0.0034446765203028917, + 0.006865826901048422, + 0.006363341584801674, + 0.006821891758590937, + 0.0077772848308086395, + 0.00620051845908165, + 0.007812262512743473, + 0.006582733243703842, + 0.014084570109844208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.038088809698820114 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.029785824939608574 + ], + "opposite_logprob_tokens": [ + -3.5297858715057373 + ], + "logprob_reward": -0.029785824939608574, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785824939608574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09046626091003418, + 0.07032458484172821, + 0.05232095718383789, + 0.13178987801074982, + 0.34098124504089355, + 0.06100448966026306, + 0.12905587255954742, + 0.02552780508995056, + 0.05510498210787773, + 0.0434238463640213 + ], + "credit_map": [ + 0.0168321393430233, + 0.00664942292496562, + 0.014742263592779636, + 0.011511857621371746, + 0.004439202137291431, + 0.010747403837740421, + 0.009880764409899712, + 0.015663210302591324, + 0.009495665319263935, + 0.006492940243333578, + 0.006959814112633467, + 0.008739039301872253, + 0.009089970029890537, + 0.007785859517753124, + 0.006400372367352247, + 0.01536092534661293, + 0.005285190884023905, + 0.005048376973718405, + 0.004585330840200186, + 0.0089446185156703, + 0.005291882436722517, + 0.004791928920894861, + 0.003321245778352022, + 0.015052380971610546, + 0.012750329449772835, + 0.025447087362408638, + 0.008483218029141426, + 0.025328729301691055, + 0.020512951537966728, + 0.01925649866461754, + 0.009396928362548351, + 0.010614139027893543, + 0.014184954576194286, + 0.013017759658396244, + 0.014769620262086391, + 0.03936503082513809, + 0.015164813958108425, + 0.13118930160999298, + 0.02188020944595337, + 0.09140955656766891, + 0.0051568797789514065, + 0.00784958153963089, + 0.0059968591667711735, + 0.008924059569835663, + 0.006528662983328104, + 0.005584286525845528, + 0.009645488113164902, + 0.011318672448396683, + 0.011138512752950191, + 0.011269999668002129, + 0.013762732036411762, + 0.013977916911244392, + 0.012909278273582458, + 0.027796339243650436, + 0.02537727914750576, + 0.012823827564716339, + 0.0025872692931443453, + 0.0019660901743918657, + 0.001953505678102374, + 0.001930602709762752, + 0.003307325066998601, + 0.0045844935812056065, + 0.00623311847448349, + 0.0029654004611074924, + 0.00485843513160944, + 0.004699876997619867, + 0.005350563675165176, + 0.004665637854486704, + 0.0049159228801727295, + 0.016389479860663414, + 0.008271248079836369, + 0.005953819025307894, + 0.004748370498418808, + 0.005707550328224897, + 0.0049155415035784245, + 0.005992583930492401, + 0.005239754915237427, + 0.005483007989823818, + 0.006418835837393999, + 0.004918200429528952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.029785824939608574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly creative first-person point-of-view shot, shot from inside the gaping mouth of a Predator warhound, its jagged, rounded teeth providing a natural frame for the frame, while its pink gums and the occasional wet tongue flickering across the camera add chaotic detail. The camera focuses on an equally goofy centaur, grappling with the warhound in a comical struggle. The centaur, sporting an exaggerated, comical expression of both panic and excitement, struggles to free itself by pushing against the warhound's face with both arms. In a childishly animated aesthetic, both characters are rendered with bold black outlines and vibrant, flat colors, their movements exaggerated and elastic. The camera shakes violently and chaotically as the warhound's head sways and the two strugglers wrestle, perfectly capturing the dynamics of the melee. The backdrop is a minimalist, brightly lit landscape of grass and blue sky, with even lighting emphasizing the scene's stylized quality. Ultimately, this scuffle, recorded from a unique perspective, creates an absurd, chaotic, and childlike atmosphere through its unstable images and the characters' clumsy interactions.", + "seed": 6657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6657_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": -0.02023784867487848, + "vlm_logprob": -0.02023784867487848, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02023784867487848, + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.035075075924396515 + ], + "opposite_logprob_tokens": [ + -3.5350751876831055 + ], + "logprob_reward": -0.035075075924396515, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035075075924396515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24178548157215118, + 0.07705177366733551, + 0.0315222404897213, + 0.05348705127835274, + 0.2799336612224579, + 0.06978770345449448, + 0.05356378108263016, + 0.07630947232246399, + 0.04625229164958, + 0.07030656933784485 + ], + "credit_map": [ + 0.038362033665180206, + 0.04173684120178223, + 0.03375014662742615, + 0.03762703761458397, + 0.019562577828764915, + 0.016988065093755722, + 0.03172120824456215, + 0.02203756757080555, + 0.014753579162061214, + 0.004942459985613823, + 0.005646123550832272, + 0.007474255748093128, + 0.005327112041413784, + 0.007988935336470604, + 0.005089201033115387, + 0.025830108672380447, + 0.004930227063596249, + 0.004571862518787384, + 0.0027141498867422342, + 0.0039058367256075144, + 0.003016081638634205, + 0.003670817706733942, + 0.003948664292693138, + 0.0047646015882492065, + 0.003756171790882945, + 0.0055516986176371574, + 0.006577549502253532, + 0.005913424771279097, + 0.008420504629611969, + 0.004801631905138493, + 0.008325258269906044, + 0.010140807367861271, + 0.023996982723474503, + 0.032847754657268524, + 0.04108859598636627, + 0.033072199672460556, + 0.049076344817876816, + 0.031592510640621185, + 0.03178539127111435, + 0.036473896354436874, + 0.011323682963848114, + 0.009345700033009052, + 0.01586681418120861, + 0.007904641330242157, + 0.0096983527764678, + 0.0041333772242069244, + 0.0059864274226129055, + 0.005528711248189211, + 0.005924468860030174, + 0.006930867210030556, + 0.006417617201805115, + 0.010514202527701855, + 0.0066133104264736176, + 0.003729586023837328, + 0.0064375633373856544, + 0.0069961645640432835, + 0.008285731077194214, + 0.010849067941308022, + 0.007972698658704758, + 0.012990294955670834, + 0.007003729231655598, + 0.006576204672455788, + 0.010195807553827763, + 0.012435932643711567, + 0.0043342686258256435, + 0.007879105396568775, + 0.004085236694663763, + 0.007658650167286396, + 0.006139022298157215, + 0.007502174004912376, + 0.004708394408226013, + 0.003945440985262394, + 0.0049565983936190605, + 0.006815716158598661, + 0.009203455410897732, + 0.013810019940137863, + 0.008799007162451744, + 0.007317907176911831, + 0.009687339887022972, + 0.009716526605188847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.035075075924396515 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.03369717672467232 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.03369717672467232, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369717672467232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16205254197120667, + 0.17427819967269897, + 0.06739792227745056, + 0.08147897571325302, + 0.18826493620872498, + 0.07482889294624329, + 0.031969886273145676, + 0.1329793483018875, + 0.05581929162144661, + 0.030929971486330032 + ], + "credit_map": [ + 0.020649362355470657, + 0.030064603313803673, + 0.02024010755121708, + 0.025540780276060104, + 0.014315212145447731, + 0.01250478532165289, + 0.02160026878118515, + 0.01713745668530464, + 0.03943926468491554, + 0.009906318970024586, + 0.012511208653450012, + 0.01674317941069603, + 0.010969015769660473, + 0.014459757134318352, + 0.010666913352906704, + 0.05958256125450134, + 0.010098961181938648, + 0.009040961042046547, + 0.005546951666474342, + 0.0070747192949056625, + 0.009637411683797836, + 0.00881348829716444, + 0.008778171613812447, + 0.00840726774185896, + 0.006726862397044897, + 0.008458620868623257, + 0.010045144706964493, + 0.008117844350636005, + 0.01697213016450405, + 0.008733890950679779, + 0.01023286022245884, + 0.012191628105938435, + 0.01632329635322094, + 0.020299427211284637, + 0.026024334132671356, + 0.020347967743873596, + 0.02971530519425869, + 0.02015955001115799, + 0.02478032559156418, + 0.030614759773015976, + 0.009497107937932014, + 0.009101121686398983, + 0.01871526427567005, + 0.010807542130351067, + 0.009300889447331429, + 0.004460457246750593, + 0.008109052665531635, + 0.004837467800825834, + 0.004207857884466648, + 0.0035501199308782816, + 0.004722340498119593, + 0.005800857674330473, + 0.0036426724400371313, + 0.0022630903404206038, + 0.0029870581347495317, + 0.004795892629772425, + 0.013455829583108425, + 0.018286656588315964, + 0.013230537995696068, + 0.023224303498864174, + 0.013766154646873474, + 0.011116879992187023, + 0.019449898973107338, + 0.020449098199605942, + 0.006534929387271404, + 0.009963887743651867, + 0.003887014929205179, + 0.008460021577775478, + 0.006835370324552059, + 0.011292674578726292, + 0.003925237338989973, + 0.004920163657516241, + 0.002462086733430624, + 0.0047719720751047134, + 0.004279266577214003, + 0.0046154106967151165, + 0.003634001361206174, + 0.003055636305361986, + 0.004427185282111168, + 0.003684415016323328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.03369717672467232 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.003192092990502715 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.003192092990502715, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192092990502715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2335721105337143, + 0.11058948934078217, + 0.07713031023740768, + 0.027525922283530235, + 0.1270177960395813, + 0.088319793343544, + 0.05280701816082001, + 0.07973408699035645, + 0.14665453135967255, + 0.05664892494678497 + ], + "credit_map": [ + 0.04064054787158966, + 0.03807211294770241, + 0.030845535919070244, + 0.026358559727668762, + 0.018918519839644432, + 0.017596814781427383, + 0.029293110594153404, + 0.03184692561626434, + 0.021558254957199097, + 0.011784669943153858, + 0.008935673162341118, + 0.01411386113613844, + 0.008702398277819157, + 0.009486574679613113, + 0.012266418896615505, + 0.023741645738482475, + 0.011603934690356255, + 0.011283555068075657, + 0.008153497241437435, + 0.009059966541826725, + 0.010728297755122185, + 0.008267237804830074, + 0.008387314155697823, + 0.009646509774029255, + 0.004176534712314606, + 0.0035781622864305973, + 0.00437697721645236, + 0.0028815590776503086, + 0.002669232664629817, + 0.002948396373540163, + 0.0029082754626870155, + 0.003986785653978586, + 0.009551319293677807, + 0.021183820441365242, + 0.014996848069131374, + 0.01621803641319275, + 0.01989687792956829, + 0.012091213837265968, + 0.02150024101138115, + 0.011579453945159912, + 0.014788375236093998, + 0.01232563890516758, + 0.009335646405816078, + 0.009465465322136879, + 0.013464870862662792, + 0.008164181374013424, + 0.006563984323292971, + 0.014211643487215042, + 0.0058242748491466045, + 0.008430532179772854, + 0.00653882697224617, + 0.009758527390658855, + 0.005066560581326485, + 0.003657340304926038, + 0.005771157797425985, + 0.0077598062343895435, + 0.011539027094841003, + 0.008239388465881348, + 0.006699458230286837, + 0.013699945993721485, + 0.008185675367712975, + 0.009378224611282349, + 0.010004650801420212, + 0.011987726204097271, + 0.0157926008105278, + 0.02771628648042679, + 0.007721958216279745, + 0.023744620382785797, + 0.014770750887691975, + 0.034696657210588455, + 0.009494894184172153, + 0.012716765515506268, + 0.004201645031571388, + 0.005987667478621006, + 0.004049534443765879, + 0.012797215953469276, + 0.006580630782991648, + 0.0043264394626021385, + 0.008560570888221264, + 0.010145226493477821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.003192092990502715 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.023282600566744804 + ], + "opposite_logprob_tokens": [ + -3.773282527923584 + ], + "logprob_reward": -0.023282600566744804, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282600566744804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13393358886241913, + 0.3421904146671295, + 0.05470466986298561, + 0.07736577838659286, + 0.09565843641757965, + 0.07850381731987, + 0.0668712705373764, + 0.05592404305934906, + 0.04869651421904564, + 0.04615150764584541 + ], + "credit_map": [ + 0.016913672909140587, + 0.017342835664749146, + 0.02126482129096985, + 0.025636350736021996, + 0.011502266861498356, + 0.008595827035605907, + 0.017461497336626053, + 0.015216303989291191, + 0.08479052037000656, + 0.024633198976516724, + 0.026041634380817413, + 0.02561686933040619, + 0.02126503922045231, + 0.02134854719042778, + 0.025566259399056435, + 0.11292833834886551, + 0.005310873035341501, + 0.009062953293323517, + 0.006702338811010122, + 0.009394998662173748, + 0.005657507106661797, + 0.004067958798259497, + 0.005344204604625702, + 0.00916383508592844, + 0.009193739853799343, + 0.009362759999930859, + 0.009698912501335144, + 0.010087673552334309, + 0.006874530576169491, + 0.008660423569381237, + 0.009819922968745232, + 0.013667814433574677, + 0.008926893584430218, + 0.012730290181934834, + 0.011256271973252296, + 0.013767208904027939, + 0.016911901533603668, + 0.009777851402759552, + 0.008915330283343792, + 0.013372679241001606, + 0.012941484339535236, + 0.010649401694536209, + 0.010703342035412788, + 0.00817797426134348, + 0.011761203408241272, + 0.010648542083799839, + 0.0062292469665408134, + 0.007392631843686104, + 0.009005588479340076, + 0.011720560491085052, + 0.008409899659454823, + 0.011733297258615494, + 0.00811601709574461, + 0.004489685408771038, + 0.0057042306289076805, + 0.0076919919811189175, + 0.006982400082051754, + 0.004736048635095358, + 0.002756771631538868, + 0.0068694171495735645, + 0.0044593410566449165, + 0.006517044268548489, + 0.006123816594481468, + 0.017479203641414642, + 0.0037434929981827736, + 0.01181611604988575, + 0.004177017603069544, + 0.007485825568437576, + 0.004706915467977524, + 0.008365306071937084, + 0.003921667579561472, + 0.004480174742639065, + 0.004554926883429289, + 0.005717671476304531, + 0.004602357745170593, + 0.008570649661123753, + 0.005963015370070934, + 0.003970160614699125, + 0.0061712670139968395, + 0.006601458881050348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.023282600566744804 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.005942297168076038 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942297168076038, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942297168076038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16681355237960815, + 0.14214058220386505, + 0.026932889595627785, + 0.06323543190956116, + 0.16942815482616425, + 0.09915804862976074, + 0.04526493698358536, + 0.07554653286933899, + 0.07305283844470978, + 0.1384269893169403 + ], + "credit_map": [ + 0.028107304126024246, + 0.028721554204821587, + 0.02327408641576767, + 0.022330142557621002, + 0.01350462343543768, + 0.011791719123721123, + 0.021494101732969284, + 0.017590034753084183, + 0.022217927500605583, + 0.010633070953190327, + 0.011229137890040874, + 0.01306675560772419, + 0.013107354752719402, + 0.022449668496847153, + 0.011343289166688919, + 0.03809337690472603, + 0.003532532136887312, + 0.003468303009867668, + 0.00242798263207078, + 0.004612843040376902, + 0.0034306240268051624, + 0.00278461747802794, + 0.002565337810665369, + 0.004110650159418583, + 0.007532993331551552, + 0.006915008649230003, + 0.011535133235156536, + 0.007240991573780775, + 0.006904432084411383, + 0.005811330862343311, + 0.0071355970576405525, + 0.01015994418412447, + 0.015241273678839207, + 0.019913310185074806, + 0.014164399355649948, + 0.012674601748585701, + 0.02910662814974785, + 0.018909016624093056, + 0.02188049629330635, + 0.037538427859544754, + 0.009269176051020622, + 0.010166653431952, + 0.011537689715623856, + 0.011349196545779705, + 0.012773985974490643, + 0.010020853951573372, + 0.009656871668994427, + 0.024383626878261566, + 0.006831417791545391, + 0.007070737425237894, + 0.005902266129851341, + 0.007069284562021494, + 0.0058116125874221325, + 0.0032892136368900537, + 0.003538725199177861, + 0.005751682911068201, + 0.007416758686304092, + 0.011022872291505337, + 0.009589679539203644, + 0.015763407573103905, + 0.006652103736996651, + 0.006642749533057213, + 0.010554803535342216, + 0.007904156111180782, + 0.006900093983858824, + 0.009617689996957779, + 0.006952621508389711, + 0.010248408652842045, + 0.011142350733280182, + 0.008290668949484825, + 0.011436127126216888, + 0.008464876562356949, + 0.010240482166409492, + 0.008617113344371319, + 0.01528992410749197, + 0.014516903087496758, + 0.015903832390904427, + 0.014347560703754425, + 0.0198602806776762, + 0.03965090215206146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005942297168076038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly creative first-person point-of-view shot, shot from inside the gaping mouth of a Predator warhound, its jagged, rounded teeth providing a natural frame for the frame, while its pink gums and the occasional wet tongue flickering across the camera add chaotic detail. The camera focuses on an equally goofy centaur, grappling with the warhound in a comical struggle. The centaur, sporting an exaggerated, comical expression of both panic and excitement, struggles to free itself by pushing against the warhound's face with both arms. In a childishly animated aesthetic, both characters are rendered with bold black outlines and vibrant, flat colors, their movements exaggerated and elastic. The camera shakes violently and chaotically as the warhound's head sways and the two strugglers wrestle, perfectly capturing the dynamics of the melee. The backdrop is a minimalist, brightly lit landscape of grass and blue sky, with even lighting emphasizing the scene's stylized quality. Ultimately, this scuffle, recorded from a unique perspective, creates an absurd, chaotic, and childlike atmosphere through its unstable images and the characters' clumsy interactions.", + "seed": 6658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_10/index_7129_seed_6658_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": -0.03792509250342846, + "vlm_logprob": -0.03792509250342846, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.03792509250342846, + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.030035266652703285 + ], + "opposite_logprob_tokens": [ + -3.7800352573394775 + ], + "logprob_reward": -0.030035266652703285, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030035266652703285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22701972723007202, + 0.054018329828977585, + 0.04367951676249504, + 0.11198115348815918, + 0.10037863999605179, + 0.137800931930542, + 0.15248221158981323, + 0.06097877398133278, + 0.03066369891166687, + 0.08099697530269623 + ], + "credit_map": [ + 0.02564752846956253, + 0.01447413582354784, + 0.06754054874181747, + 0.03568858280777931, + 0.017269104719161987, + 0.02120259776711464, + 0.02703649364411831, + 0.018160730600357056, + 0.006099822465330362, + 0.006466206628829241, + 0.005343006923794746, + 0.012156562879681587, + 0.008709389716386795, + 0.0062390039674937725, + 0.005248102825134993, + 0.0037562320940196514, + 0.0037631166633218527, + 0.005383501760661602, + 0.004676739685237408, + 0.00564891891553998, + 0.006129265762865543, + 0.00994849018752575, + 0.005517900455743074, + 0.0026115821674466133, + 0.01220092922449112, + 0.014263661578297615, + 0.012092878110706806, + 0.022555261850357056, + 0.009342777542769909, + 0.014167014509439468, + 0.013853526674211025, + 0.013505106791853905, + 0.011090664193034172, + 0.0073234038427472115, + 0.011814131401479244, + 0.011764440685510635, + 0.011027289554476738, + 0.022456003352999687, + 0.012089187279343605, + 0.012813526205718517, + 0.011331714689731598, + 0.012049240991473198, + 0.0090075908228755, + 0.012986093759536743, + 0.01968485675752163, + 0.023951269686222076, + 0.02995225600898266, + 0.018837913870811462, + 0.019765278324484825, + 0.015608880668878555, + 0.021852809935808182, + 0.036629341542720795, + 0.011602831073105335, + 0.015510008670389652, + 0.018123924732208252, + 0.013389138504862785, + 0.00979938730597496, + 0.013391564600169659, + 0.006516304798424244, + 0.0054752035066485405, + 0.006602645851671696, + 0.005751427263021469, + 0.008847255259752274, + 0.004594982136040926, + 0.00414640735834837, + 0.003092491300776601, + 0.0023343891371041536, + 0.0025489297695457935, + 0.003805828280746937, + 0.002640501596033573, + 0.008021760731935501, + 0.004073391202837229, + 0.009565578773617744, + 0.008929598145186901, + 0.011556310579180717, + 0.01014803722500801, + 0.014181945472955704, + 0.00683084549382329, + 0.009602311998605728, + 0.010182344354689121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.030035266652703285 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.04306533932685852 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306533932685852, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306533932685852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14567780494689941, + 0.024735011160373688, + 0.3472347557544708, + 0.06319809705018997, + 0.06844687461853027, + 0.21205571293830872, + 0.027467219159007072, + 0.04305894300341606, + 0.014018992893397808, + 0.05410662665963173 + ], + "credit_map": [ + 0.020013900473713875, + 0.007838737219572067, + 0.04301171749830246, + 0.02738001011312008, + 0.00642568850889802, + 0.011948429979383945, + 0.012105721980333328, + 0.016953587532043457, + 0.002776358975097537, + 0.0034414492547512054, + 0.002377081196755171, + 0.003438435262069106, + 0.0032472233287990093, + 0.0033694361336529255, + 0.003917515277862549, + 0.002167511498555541, + 0.04134970158338547, + 0.016043605282902718, + 0.009828654117882252, + 0.009484226815402508, + 0.0719439834356308, + 0.15484298765659332, + 0.028253300115466118, + 0.015488275326788425, + 0.008536851033568382, + 0.012097608298063278, + 0.009163971059024334, + 0.005245478358119726, + 0.00787054281681776, + 0.008758721873164177, + 0.0060881380923092365, + 0.0054367841221392155, + 0.007614994887262583, + 0.005828883498907089, + 0.011469251476228237, + 0.009325588122010231, + 0.0076701161451637745, + 0.011709431186318398, + 0.006366773974150419, + 0.008461837656795979, + 0.011903326027095318, + 0.011806623078882694, + 0.005342008080333471, + 0.007446439005434513, + 0.014504075050354004, + 0.07103993743658066, + 0.07502859085798264, + 0.014984725043177605, + 0.00511014461517334, + 0.0036987310741096735, + 0.00367755233310163, + 0.0038300410378724337, + 0.003101862734183669, + 0.0027520961593836546, + 0.003355292370542884, + 0.0019414990674704313, + 0.006287775002419949, + 0.007749464828521013, + 0.006327097769826651, + 0.005248578265309334, + 0.004392685834318399, + 0.003969939425587654, + 0.005495037417858839, + 0.0035883677192032337, + 0.0011080673430114985, + 0.0011132434010505676, + 0.0013562622480094433, + 0.00107974570710212, + 0.0018619311740621924, + 0.0013067637337371707, + 0.003912047948688269, + 0.002280931221321225, + 0.009693941101431847, + 0.006699457764625549, + 0.003111806232482195, + 0.002944792155176401, + 0.004378427751362324, + 0.019035911187529564, + 0.0033877950627356768, + 0.004854493774473667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.04306533932685852 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.029789064079523087 + ], + "opposite_logprob_tokens": [ + -3.5297889709472656 + ], + "logprob_reward": -0.029789064079523087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789064079523087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25821229815483093, + 0.036629818379879, + 0.40428560972213745, + 0.03181714937090874, + 0.05018499121069908, + 0.046502482146024704, + 0.05092963948845863, + 0.02274729125201702, + 0.05255543440580368, + 0.04613533988595009 + ], + "credit_map": [ + 0.025623975321650505, + 0.01614171452820301, + 0.021077483892440796, + 0.0838860496878624, + 0.015155777335166931, + 0.014132407493889332, + 0.061103455722332, + 0.021091405302286148, + 0.003791020018979907, + 0.0049751149490475655, + 0.003054946893826127, + 0.00760290864855051, + 0.004105615429580212, + 0.004667568951845169, + 0.005501541309058666, + 0.0029310989193618298, + 0.04764605686068535, + 0.019122129306197166, + 0.011461556889116764, + 0.011895538307726383, + 0.08489596098661423, + 0.180097758769989, + 0.03284348547458649, + 0.016323089599609375, + 0.004782753065228462, + 0.0031348306220024824, + 0.003695988329127431, + 0.004453607369214296, + 0.003659279318526387, + 0.004012044053524733, + 0.004042796324938536, + 0.0040358491241931915, + 0.0058755348436534405, + 0.005801843013614416, + 0.005801083520054817, + 0.007081836927682161, + 0.005338585469871759, + 0.006533756852149963, + 0.006696724332869053, + 0.007055624388158321, + 0.0032463695388287306, + 0.0030199983157217503, + 0.0023809620179235935, + 0.0035337090957909822, + 0.006876198574900627, + 0.010973429307341576, + 0.011835923418402672, + 0.004635891877114773, + 0.006384690757840872, + 0.006805545650422573, + 0.007625897414982319, + 0.007783324923366308, + 0.005214518867433071, + 0.005558238830417395, + 0.007432177662849426, + 0.004125244449824095, + 0.003755672136321664, + 0.0033093919046223164, + 0.0026767796371132135, + 0.0030122101306915283, + 0.002173137851059437, + 0.00214031501673162, + 0.0031222549732774496, + 0.0025575277395546436, + 0.007019944489002228, + 0.0071821920573711395, + 0.003529902547597885, + 0.0042998599819839, + 0.00685072923079133, + 0.004385201260447502, + 0.01271346490830183, + 0.006574137602001429, + 0.00798924919217825, + 0.0057476600632071495, + 0.004657462704926729, + 0.003454597434028983, + 0.003979461267590523, + 0.0144547950476408, + 0.0025748531334102154, + 0.003277262905612588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.029789064079523087 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.03809259459376335 + ], + "opposite_logprob_tokens": [ + -3.288092613220215 + ], + "logprob_reward": -0.03809259459376335, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809259459376335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10949771106243134, + 0.03994208201766014, + 0.35787785053253174, + 0.10239728540182114, + 0.06098062917590141, + 0.18021567165851593, + 0.0207191314548254, + 0.04693841189146042, + 0.03729579970240593, + 0.04413541406393051 + ], + "credit_map": [ + 0.008595209568738937, + 0.006778524722903967, + 0.020251452922821045, + 0.0358603410422802, + 0.0067894780077040195, + 0.007188941352069378, + 0.014595863409340382, + 0.009437896311283112, + 0.004961967468261719, + 0.005244528409093618, + 0.0036471609491854906, + 0.004868374206125736, + 0.005155998282134533, + 0.0052256654016673565, + 0.0065629733726382256, + 0.004275417886674404, + 0.04511965066194534, + 0.016873888671398163, + 0.011066838167607784, + 0.009987740777432919, + 0.06834638863801956, + 0.16072063148021698, + 0.03085716813802719, + 0.01490551233291626, + 0.013054999522864819, + 0.013729567639529705, + 0.012701433151960373, + 0.016280202195048332, + 0.013047468848526478, + 0.012269913218915462, + 0.011525311507284641, + 0.009788396768271923, + 0.007387436460703611, + 0.005349909886717796, + 0.008384214714169502, + 0.006346633192151785, + 0.004981923382729292, + 0.014928186312317848, + 0.008623919449746609, + 0.004978406708687544, + 0.009802041575312614, + 0.009359819814562798, + 0.00502739567309618, + 0.006745942402631044, + 0.011031134985387325, + 0.06443103402853012, + 0.06291551887989044, + 0.01090277824550867, + 0.0017047179862856865, + 0.0017987593309953809, + 0.002768944948911667, + 0.00250252359546721, + 0.002307229908183217, + 0.002460265066474676, + 0.004149824846535921, + 0.00302686495706439, + 0.0061452253721654415, + 0.006639741826802492, + 0.007002175785601139, + 0.005600250791758299, + 0.006650043185800314, + 0.003430674783885479, + 0.006929485592991114, + 0.0045408145524561405, + 0.00501206936314702, + 0.006611932534724474, + 0.004120683763176203, + 0.0037393192760646343, + 0.00513197248801589, + 0.003128949087113142, + 0.005179971922188997, + 0.0043709008023142815, + 0.006196302827447653, + 0.004296739585697651, + 0.009258261881768703, + 0.007700955029577017, + 0.003120809094980359, + 0.0067532360553741455, + 0.0036283668596297503, + 0.0031807434279471636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03809259459376335 + }, + { + "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 inside the gaping mouth of a Predator warhound with jagged, rounded teeth framing the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a centaur visible grappling with the warhound in a comical struggle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the characters rendered with bold black outlines and vibrant, flat colors in a childishly animated aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake violently and chaotically as the warhound's head sways and the characters wrestle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a minimalist, brightly lit landscape of grass and blue 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": [ + -0.04864319786429405 + ], + "opposite_logprob_tokens": [ + -3.048643112182617 + ], + "logprob_reward": -0.04864319786429405, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864319786429405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17475949227809906, + 0.027452602982521057, + 0.20077890157699585, + 0.06980622559785843, + 0.24304711818695068, + 0.10017641633749008, + 0.04860151931643486, + 0.08105222135782242, + 0.031446464359760284, + 0.022879043594002724 + ], + "credit_map": [ + 0.02767355367541313, + 0.012095912359654903, + 0.047227706760168076, + 0.02595934458076954, + 0.008753524161875248, + 0.01371007040143013, + 0.017259618267416954, + 0.02207975648343563, + 0.002806640462949872, + 0.004136931151151657, + 0.002452690154314041, + 0.00504054082557559, + 0.0039299242198467255, + 0.002884858986362815, + 0.003606899408623576, + 0.0025941182393580675, + 0.02462150529026985, + 0.010813586413860321, + 0.008033732883632183, + 0.007259207312017679, + 0.038086652755737305, + 0.08592741936445236, + 0.01700354740023613, + 0.009033231064677238, + 0.00906505435705185, + 0.00835166685283184, + 0.006549369543790817, + 0.008732997812330723, + 0.009836441837251186, + 0.00978259276598692, + 0.007436477579176426, + 0.010051621124148369, + 0.03605153039097786, + 0.0314045213162899, + 0.039852775633335114, + 0.025032466277480125, + 0.021612955257296562, + 0.04017949849367142, + 0.025317875668406487, + 0.02359548583626747, + 0.005305852275341749, + 0.006062718573957682, + 0.004203739110380411, + 0.0057389335706830025, + 0.00928674079477787, + 0.03201733157038689, + 0.02966456487774849, + 0.007896536029875278, + 0.006417333614081144, + 0.006085036788135767, + 0.004870113916695118, + 0.004584027919918299, + 0.00656847981736064, + 0.005186393391340971, + 0.00910639576613903, + 0.005783737171441317, + 0.01296449825167656, + 0.012824337929487228, + 0.012045335955917835, + 0.009657960385084152, + 0.008553425781428814, + 0.007308419793844223, + 0.01031419076025486, + 0.007384059019386768, + 0.004557844717055559, + 0.004000118002295494, + 0.002150657121092081, + 0.002901732921600342, + 0.004169672727584839, + 0.002725837053731084, + 0.007651452906429768, + 0.0032891505397856236, + 0.0024309006985276937, + 0.0025168494321405888, + 0.003714801976457238, + 0.002793149324133992, + 0.003347888356074691, + 0.0029262693133205175, + 0.002714075380936265, + 0.00243510864675045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.04864319786429405 + } + ], + "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/0000017/group_11/index_tensor([3263, 3263])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_tensor([3263, 3263])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bdd0c85f78486d1a43247f1eb2ec40c5f3e9f146 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_tensor([3263, 3263])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, cinematic shot begins, moving from left to right, capturing a wide shot of a bustling open-air market on a sun-drenched cobblestone square. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long, gentle shadows and creating a warm, inviting atmosphere. The camera first glides past a vendor's stall, where vibrant powdered spices—fiery red chili powder, golden turmeric, and earthy brown cumin—are piled high in large wooden bowls. Next to them, stacks of meticulously handwoven baskets made of natural straw and spandex display intricate patterns. The camera continues in a steady motion, showcasing a family gathered together; a mother, her hair tied in a loose bun and wearing a simple white linen dress, gently places a small, shimmering candy into her daughter's waiting hands. As the child savors it, her eyes widen with delight. In the background, other shoppers stroll between stalls shaded by colorful fabric awnings, their muffled conversation blending with the gentle, melodic notes of a lute played by a musician seated on a low stool. The pot finally fell on a little boy in a bright blue T-shirt, who excitedly tugged at his father's sleeve, stretched out his arm, and pointed with pure wonder at a stall overflowing with colorful, sugar-dusted candies, their surfaces gleaming in the warm light.", + "seed": 17975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_3263_seed_17975_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": -0.3976091824006289, + "vlm_logprob": -0.3976091824006289, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.3976091824006289, + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.91143798828125 + ], + "opposite_logprob_tokens": [ + -0.16143803298473358 + ], + "logprob_reward": -1.91143798828125, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.91143798828125, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16181716322898865, + 0.03856887295842171, + 0.02704435959458351, + 0.5076678395271301, + 0.08826093375682831, + 0.04368007555603981, + 0.02402048371732235, + 0.027270035818219185, + 0.04451598599553108, + 0.03715430945158005 + ], + "credit_map": [ + 0.01815040037035942, + 0.022098861634731293, + 0.022030537948012352, + 0.027042271569371223, + 0.017836719751358032, + 0.015113197267055511, + 0.022321665659546852, + 0.017223486676812172, + 0.0030791109893471003, + 0.0054049137979745865, + 0.005428689066320658, + 0.006887827068567276, + 0.0036158855073153973, + 0.003632515436038375, + 0.006704638712108135, + 0.0038152914494276047, + 0.004816716071218252, + 0.0033162254840135574, + 0.004798733163625002, + 0.0031296780798584223, + 0.0022304789163172245, + 0.0027243674267083406, + 0.0038410762790590525, + 0.0021870830096304417, + 0.035587795078754425, + 0.013815837912261486, + 0.12775743007659912, + 0.04790697619318962, + 0.01810998097062111, + 0.029754677787423134, + 0.06560944020748138, + 0.16912566125392914, + 0.012869011610746384, + 0.010862631723284721, + 0.018528204411268234, + 0.01097468938678503, + 0.008416726253926754, + 0.006728845648467541, + 0.01259615272283554, + 0.007284666411578655, + 0.008179159834980965, + 0.006333362776786089, + 0.0058448235504329205, + 0.006089629139751196, + 0.0027305507101118565, + 0.0032761201728135347, + 0.007560759782791138, + 0.0036656660959124565, + 0.003918695263564587, + 0.003467411734163761, + 0.0035932203754782677, + 0.003260938683524728, + 0.0022775009274482727, + 0.0019335554679855704, + 0.00312783382833004, + 0.002441328251734376, + 0.004853955935686827, + 0.0024570762179791927, + 0.005497688427567482, + 0.003426031209528446, + 0.0025068854447454214, + 0.002623563399538398, + 0.003228088840842247, + 0.00267674308270216, + 0.005311616696417332, + 0.003843765938654542, + 0.0024783527478575706, + 0.005819347687065601, + 0.0050798989832401276, + 0.008591105230152607, + 0.007916617207229137, + 0.00547527801245451, + 0.00135481683537364, + 0.0014209975488483906, + 0.0012953992700204253, + 0.0020929588936269283, + 0.001834988477639854, + 0.002807802055031061, + 0.01465192437171936, + 0.011695417575538158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -1.91143798828125 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03819288685917854 + ], + "opposite_logprob_tokens": [ + -3.2881929874420166 + ], + "logprob_reward": -0.03819288685917854, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03819288685917854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1168152242898941, + 0.04098827391862869, + 0.02644314616918564, + 0.5702071189880371, + 0.07066508382558823, + 0.052938275039196014, + 0.021337473765015602, + 0.0497983917593956, + 0.03731578588485718, + 0.013491150923073292 + ], + "credit_map": [ + 0.01384038757532835, + 0.01597006991505623, + 0.017180996015667915, + 0.019422417506575584, + 0.012567958794534206, + 0.010595218278467655, + 0.015174875035881996, + 0.012063306756317616, + 0.0035740204621106386, + 0.006533951032906771, + 0.005525875836610794, + 0.007731550373136997, + 0.004208757542073727, + 0.003915595356374979, + 0.005516089498996735, + 0.0039824373088777065, + 0.0047141872346401215, + 0.0036178056616336107, + 0.0051277028396725655, + 0.0024880431592464447, + 0.002044767141342163, + 0.0027082364540547132, + 0.003633670974522829, + 0.002108732471242547, + 0.04031116142868996, + 0.016833432018756866, + 0.14775793254375458, + 0.05508030205965042, + 0.020296907052397728, + 0.0335146002471447, + 0.06798847764730453, + 0.18842439353466034, + 0.010399636812508106, + 0.00862153060734272, + 0.014992156997323036, + 0.008805868215858936, + 0.006835121661424637, + 0.0050231595523655415, + 0.01013383362442255, + 0.005853779148310423, + 0.008721977472305298, + 0.007800684310495853, + 0.006846119649708271, + 0.007592478767037392, + 0.003627379424870014, + 0.003869449719786644, + 0.010397077538073063, + 0.004083109088242054, + 0.0035225406754761934, + 0.0035866629332304, + 0.003132015699520707, + 0.00276365433819592, + 0.002154786139726639, + 0.0017666901694610715, + 0.0022935031447559595, + 0.0021176221780478954, + 0.008413825184106827, + 0.004689223598688841, + 0.00906476192176342, + 0.005697737447917461, + 0.0038902517408132553, + 0.004781976342201233, + 0.007487704511731863, + 0.0057729147374629974, + 0.0038496055640280247, + 0.0033668535761535168, + 0.0020881537348031998, + 0.004845593590289354, + 0.004363341256976128, + 0.006898561492562294, + 0.00713764876127243, + 0.0047660283744335175, + 0.0014617551350966096, + 0.0015191161073744297, + 0.0015645988751202822, + 0.0019449981627985835, + 0.0011731103295460343, + 0.0018949760124087334, + 0.0023073048796504736, + 0.0016252929344773293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.03819288685917854 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628576872870326 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.003628576872870326, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628576872870326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1493077278137207, + 0.058411892503499985, + 0.025573283433914185, + 0.5170422196388245, + 0.08236772567033768, + 0.031460635364055634, + 0.022365868091583252, + 0.023971425369381905, + 0.055854059755802155, + 0.03364524245262146 + ], + "credit_map": [ + 0.017137620598077774, + 0.020012933760881424, + 0.01964545249938965, + 0.02410939894616604, + 0.01578625477850437, + 0.014982272870838642, + 0.020854424685239792, + 0.0167793370783329, + 0.004790755920112133, + 0.00925695151090622, + 0.0075294747948646545, + 0.011455695144832134, + 0.005461538210511208, + 0.005199848208576441, + 0.010307779535651207, + 0.00440983846783638, + 0.005275293719023466, + 0.0033954395912587643, + 0.004727136809378862, + 0.002267559990286827, + 0.002083495259284973, + 0.002572267549112439, + 0.003221658291295171, + 0.002030429197475314, + 0.036045003682374954, + 0.01428029965609312, + 0.1295304000377655, + 0.04875761643052101, + 0.018447285518050194, + 0.030226193368434906, + 0.06766216456890106, + 0.1720932275056839, + 0.012093714438378811, + 0.010070348158478737, + 0.017046667635440826, + 0.01038484089076519, + 0.007865983061492443, + 0.006328441668301821, + 0.011648792773485184, + 0.006928928662091494, + 0.0046915533021092415, + 0.004553680773824453, + 0.0039495727978646755, + 0.004790111910551786, + 0.0020620625000447035, + 0.0021312085445970297, + 0.0060059125535190105, + 0.003276530420407653, + 0.003628564765676856, + 0.00333228032104671, + 0.003348113503307104, + 0.0028428344521671534, + 0.002238991903141141, + 0.0017654953990131617, + 0.0029289531521499157, + 0.0022806336637586355, + 0.004534950479865074, + 0.0020654918625950813, + 0.004783024080097675, + 0.002930683083832264, + 0.0023771882988512516, + 0.0022205132991075516, + 0.0026650794316083193, + 0.0023944913409650326, + 0.0071571883745491505, + 0.004381990991532803, + 0.0047637769021093845, + 0.0065295095555484295, + 0.005816415883600712, + 0.013931402005255222, + 0.007062004879117012, + 0.006211764179170132, + 0.0011385191464796662, + 0.001573608024045825, + 0.001410024706274271, + 0.0021381753031164408, + 0.001940815825946629, + 0.003189363982528448, + 0.012297260574996471, + 0.009957472793757915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.003628576872870326 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594803616404533 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594803616404533, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594803616404533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09525871276855469, + 0.04176170751452446, + 0.022006114944815636, + 0.2981208562850952, + 0.06388279795646667, + 0.03412124887108803, + 0.022417880594730377, + 0.02051597647368908, + 0.044471610337495804, + 0.35744309425354004 + ], + "credit_map": [ + 0.011504496447741985, + 0.012496045790612698, + 0.012800350785255432, + 0.016838740557432175, + 0.009556200355291367, + 0.008978810161352158, + 0.013449263758957386, + 0.00963479932397604, + 0.0034120192285627127, + 0.006682632490992546, + 0.005356983281672001, + 0.008796061389148235, + 0.003716528881341219, + 0.0037208180874586105, + 0.006792873609811068, + 0.003283788450062275, + 0.0038304137997329235, + 0.002474919892847538, + 0.0035902985837310553, + 0.0029617585241794586, + 0.0018937572604045272, + 0.002052454510703683, + 0.0033823561388999224, + 0.0018201543716713786, + 0.021325476467609406, + 0.008970323950052261, + 0.07523103803396225, + 0.028121013194322586, + 0.010646634735167027, + 0.017381025478243828, + 0.037755679339170456, + 0.09868963807821274, + 0.009485773742198944, + 0.00831638090312481, + 0.013311104848980904, + 0.007853529416024685, + 0.005742325447499752, + 0.00432121567428112, + 0.009302997961640358, + 0.005549464840441942, + 0.004487826023250818, + 0.004624828230589628, + 0.004102320410311222, + 0.0043871477246284485, + 0.003431762335821986, + 0.0029735127463936806, + 0.005606047809123993, + 0.004507800098508596, + 0.002079938305541873, + 0.0027264622040092945, + 0.0035859181080013514, + 0.0027814952190965414, + 0.003373424056917429, + 0.002683883998543024, + 0.0032294956035912037, + 0.0019572614692151546, + 0.002467856276780367, + 0.002283605048432946, + 0.0029535815119743347, + 0.0019967376720160246, + 0.0028491434641182423, + 0.002617449965327978, + 0.003088087774813175, + 0.0022595133632421494, + 0.004590390715748072, + 0.00514189014211297, + 0.002326269168406725, + 0.005184403154999018, + 0.006252387538552284, + 0.007298119366168976, + 0.007803070824593306, + 0.005875074304640293, + 0.011175928637385368, + 0.009410226717591286, + 0.007556815166026354, + 0.014205327257514, + 0.014984267763793468, + 0.018479062244296074, + 0.15563298761844635, + 0.12599845230579376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.020594803616404533 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1256030946969986, + 0.0427606962621212, + 0.024716079235076904, + 0.3960660398006439, + 0.06864643096923828, + 0.03504064679145813, + 0.029607834294438362, + 0.05886602774262428, + 0.08845044672489166, + 0.130242720246315 + ], + "credit_map": [ + 0.01471023540943861, + 0.0170830599963665, + 0.016352206468582153, + 0.022327329963445663, + 0.013740510679781437, + 0.011203352361917496, + 0.0174079779535532, + 0.01277842279523611, + 0.0032014690805226564, + 0.006042946130037308, + 0.006026968825608492, + 0.008074005134403706, + 0.003483804874122143, + 0.0038064545951783657, + 0.007503953762352467, + 0.004621093161404133, + 0.0035898853093385696, + 0.003541427431628108, + 0.00418747728690505, + 0.0030853503849357367, + 0.002423136029392481, + 0.0024449490010738373, + 0.003311891807243228, + 0.0021319594234228134, + 0.027599511668086052, + 0.012345477938652039, + 0.09820732474327087, + 0.03701996058225632, + 0.014491934329271317, + 0.022930480539798737, + 0.0521499402821064, + 0.13132138550281525, + 0.010086655616760254, + 0.008872230537235737, + 0.014982662163674831, + 0.00801657885313034, + 0.006746760569512844, + 0.0047630323097109795, + 0.00916988030076027, + 0.006008628290146589, + 0.00400331849232316, + 0.005796761717647314, + 0.004231438506394625, + 0.0053594415076076984, + 0.0025872939731925726, + 0.0024404271971434355, + 0.006836936343461275, + 0.0037850283551961184, + 0.004115263931453228, + 0.003705497132614255, + 0.004545421805232763, + 0.004198969807475805, + 0.0025492345448583364, + 0.002835841616615653, + 0.004345158115029335, + 0.0033124464098364115, + 0.008668176829814911, + 0.006272253580391407, + 0.008894464001059532, + 0.007761859800666571, + 0.007934849709272385, + 0.005877315998077393, + 0.007755482103675604, + 0.005701627116650343, + 0.011884481646120548, + 0.009029711596667767, + 0.01032226625829935, + 0.012641808949410915, + 0.008476479910314083, + 0.017934301868081093, + 0.009107491001486778, + 0.009053907357156277, + 0.005236521363258362, + 0.006477725692093372, + 0.005505966022610664, + 0.007856854237616062, + 0.009246899746358395, + 0.011523911729454994, + 0.04754282161593437, + 0.036852020770311356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.01419165637344122 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic shot begins, moving from left to right, capturing a wide shot of a bustling open-air market on a sun-drenched cobblestone square. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long, gentle shadows and creating a warm, inviting atmosphere. The camera first glides past a vendor's stall, where vibrant powdered spices—fiery red chili powder, golden turmeric, and earthy brown cumin—are piled high in large wooden bowls. Next to them, stacks of meticulously handwoven baskets made of natural straw and spandex display intricate patterns. The camera continues in a steady motion, showcasing a family gathered together; a mother, her hair tied in a loose bun and wearing a simple white linen dress, gently places a small, shimmering candy into her daughter's waiting hands. As the child savors it, her eyes widen with delight. In the background, other shoppers stroll between stalls shaded by colorful fabric awnings, their muffled conversation blending with the gentle, melodic notes of a lute played by a musician seated on a low stool. The pot finally fell on a little boy in a bright blue T-shirt, who excitedly tugged at his father's sleeve, stretched out his arm, and pointed with pure wonder at a stall overflowing with colorful, sugar-dusted candies, their surfaces gleaming in the warm light.", + "seed": 17976, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_3263_seed_17976_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": -0.14569979370571673, + "vlm_logprob": -0.14569979370571673, + "vlm_yes_no_margin": 2.8499999761581423, + "vlm_token_credit": -0.14569979370571673, + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15329228341579437 + ], + "opposite_logprob_tokens": [ + -2.028292179107666 + ], + "logprob_reward": -0.15329228341579437, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15329228341579437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09012247622013092, + 0.1777043342590332, + 0.09443403780460358, + 0.03754270821809769, + 0.10463657975196838, + 0.21604688465595245, + 0.06782799959182739, + 0.031151803210377693, + 0.10974892228841782, + 0.07078429311513901 + ], + "credit_map": [ + 0.010338257066905499, + 0.013947736471891403, + 0.011130235157907009, + 0.011856782250106335, + 0.008078740909695625, + 0.012320921756327152, + 0.01089500356465578, + 0.011554790660738945, + 0.022462833672761917, + 0.01378768403083086, + 0.023067135363817215, + 0.019932957366108894, + 0.018317675217986107, + 0.025761332362890244, + 0.030061952769756317, + 0.02431277371942997, + 0.013415270484983921, + 0.013543671928346157, + 0.009713156148791313, + 0.016863763332366943, + 0.00874652061611414, + 0.008976844139397144, + 0.01423744298517704, + 0.008937360718846321, + 0.005114985629916191, + 0.002909369533881545, + 0.005262711551040411, + 0.007157495245337486, + 0.0032464107498526573, + 0.003432195633649826, + 0.004683404695242643, + 0.005736134480684996, + 0.009479060769081116, + 0.007865955121815205, + 0.01440521702170372, + 0.02270585112273693, + 0.008965568616986275, + 0.008927855640649796, + 0.018681881949305534, + 0.013605180196464062, + 0.009876915253698826, + 0.00584655674174428, + 0.010865547694265842, + 0.02260945551097393, + 0.014756793156266212, + 0.011546149849891663, + 0.08171069622039795, + 0.058834753930568695, + 0.00673609459772706, + 0.007412616163492203, + 0.007592565380036831, + 0.00787094421684742, + 0.008289134129881859, + 0.009866840206086636, + 0.010445238091051579, + 0.009614570066332817, + 0.0028615163173526525, + 0.0026882446836680174, + 0.004401231184601784, + 0.003916027024388313, + 0.00418276758864522, + 0.0034964652732014656, + 0.005762648303061724, + 0.003842903533950448, + 0.01171813253313303, + 0.00819830596446991, + 0.016917137429118156, + 0.025464994832873344, + 0.012076448649168015, + 0.011854550801217556, + 0.012294478714466095, + 0.011224870570003986, + 0.0067026750184595585, + 0.008291346952319145, + 0.006325394380837679, + 0.008540458045899868, + 0.012093598023056984, + 0.010431571863591671, + 0.01007931400090456, + 0.008319932967424393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.15329228341579437 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497979745268822 + ], + "opposite_logprob_tokens": [ + -2.9299798011779785 + ], + "logprob_reward": -0.05497979745268822, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497979745268822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08052000403404236, + 0.21692942082881927, + 0.06882135570049286, + 0.05363960191607475, + 0.1208374872803688, + 0.18785609304904938, + 0.12933124601840973, + 0.019119707867503166, + 0.06784329563379288, + 0.05510185658931732 + ], + "credit_map": [ + 0.006636695470660925, + 0.007631361950188875, + 0.008949585258960724, + 0.019315553829073906, + 0.00629652664065361, + 0.009153611026704311, + 0.011217162944376469, + 0.011319497600197792, + 0.024295613169670105, + 0.02535678632557392, + 0.03598892316222191, + 0.025410644710063934, + 0.020326919853687286, + 0.033810585737228394, + 0.032098088413476944, + 0.019641852006316185, + 0.008152936585247517, + 0.008086205460131168, + 0.010591165162622929, + 0.009684089571237564, + 0.006003024522215128, + 0.009610726498067379, + 0.008810210041701794, + 0.007882996462285519, + 0.007339221425354481, + 0.0046575479209423065, + 0.008171163499355316, + 0.011500613763928413, + 0.0043498375453054905, + 0.005527667235583067, + 0.006561229936778545, + 0.005532316863536835, + 0.015903130173683167, + 0.008673924021422863, + 0.014979607425630093, + 0.02729521133005619, + 0.011359813623130322, + 0.011785115115344524, + 0.01870395615696907, + 0.012136726640164852, + 0.02253207005560398, + 0.01800682209432125, + 0.018351592123508453, + 0.014867188408970833, + 0.02275419421494007, + 0.021343158558011055, + 0.036213263869285583, + 0.03378778696060181, + 0.01039204653352499, + 0.008991412818431854, + 0.017193535342812538, + 0.020993266254663467, + 0.015552322380244732, + 0.01710912212729454, + 0.020915796980261803, + 0.01818372868001461, + 0.0014095711521804333, + 0.001779085025191307, + 0.0025185018312186003, + 0.00314507563598454, + 0.0022312160581350327, + 0.0022824471816420555, + 0.0026905362028628588, + 0.003063274547457695, + 0.006517506204545498, + 0.007988198660314083, + 0.006485080346465111, + 0.010477217845618725, + 0.008532957173883915, + 0.006451859604567289, + 0.014195741154253483, + 0.0071947285905480385, + 0.004089398309588432, + 0.0055671026930212975, + 0.006931959185749292, + 0.01239704992622137, + 0.005451136734336615, + 0.006244356743991375, + 0.008383299224078655, + 0.006037552375346422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.05497979745268822 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002830310957506299 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.002830310957506299, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002830310957506299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2530424892902374, + 0.0764901340007782, + 0.2082434892654419, + 0.04388817399740219, + 0.06517715007066727, + 0.11811333894729614, + 0.06577801704406738, + 0.039103616029024124, + 0.05434096232056618, + 0.0758226066827774 + ], + "credit_map": [ + 0.02835654467344284, + 0.050250615924596786, + 0.028058074414730072, + 0.03399733453989029, + 0.023386863991618156, + 0.031612277030944824, + 0.021553346887230873, + 0.03582741320133209, + 0.011012649163603783, + 0.008981884457170963, + 0.009772018529474735, + 0.009902349673211575, + 0.005973251536488533, + 0.010096296668052673, + 0.013630317524075508, + 0.0071213701739907265, + 0.03403119742870331, + 0.027411993592977524, + 0.03252621740102768, + 0.029760517179965973, + 0.015384750440716743, + 0.01480916142463684, + 0.029918834567070007, + 0.02440081350505352, + 0.008831302635371685, + 0.003541972255334258, + 0.007108522113412619, + 0.00851654913276434, + 0.003625283483415842, + 0.0031910252291709185, + 0.004165069665759802, + 0.0049084508791565895, + 0.004980148281902075, + 0.006044104695320129, + 0.010545971803367138, + 0.01452252734452486, + 0.007628203835338354, + 0.005953268148005009, + 0.00909886322915554, + 0.006404064130038023, + 0.006145849823951721, + 0.005371392238885164, + 0.008677473291754723, + 0.012064562179148197, + 0.011930341832339764, + 0.008168182335793972, + 0.04110841453075409, + 0.024647129699587822, + 0.0060376813635230064, + 0.008805554360151291, + 0.008256608620285988, + 0.010393648408353329, + 0.008908196352422237, + 0.007481679320335388, + 0.009704184718430042, + 0.006190467160195112, + 0.0026065160054713488, + 0.0035542587283998728, + 0.0045756325125694275, + 0.0055317990481853485, + 0.00503651425242424, + 0.0042405459098517895, + 0.007834549061954021, + 0.00572380144149065, + 0.004818886984139681, + 0.004490608349442482, + 0.008896730840206146, + 0.009678679518401623, + 0.0057212188839912415, + 0.005605692509561777, + 0.009145942516624928, + 0.005983204580843449, + 0.006109208799898624, + 0.007863715291023254, + 0.008221616968512535, + 0.014548608101904392, + 0.009152494370937347, + 0.009320585988461971, + 0.011651014909148216, + 0.008955364115536213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.002830310957506299 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47434425354003906 + ], + "opposite_logprob_tokens": [ + -0.9743442535400391 + ], + "logprob_reward": -0.47434425354003906, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47434425354003906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1512507051229477, + 0.11365074664354324, + 0.0694393515586853, + 0.036558836698532104, + 0.13618220388889313, + 0.15518267452716827, + 0.06473899632692337, + 0.08348827064037323, + 0.06636302918195724, + 0.12314526736736298 + ], + "credit_map": [ + 0.01593896374106407, + 0.03666212782263756, + 0.016104018315672874, + 0.013301827944815159, + 0.01456745807081461, + 0.018918361514806747, + 0.01454765722155571, + 0.021210255101323128, + 0.013309651054441929, + 0.009385461919009686, + 0.015267736278474331, + 0.013031302019953728, + 0.013249947689473629, + 0.014795768074691296, + 0.01782185770571232, + 0.016788991168141365, + 0.00865516159683466, + 0.008066710084676743, + 0.010680868290364742, + 0.009604566730558872, + 0.004523881245404482, + 0.007051190827041864, + 0.00868313666433096, + 0.012173821218311787, + 0.004728856962174177, + 0.0032976889051496983, + 0.004213529638946056, + 0.007820085622370243, + 0.0034458693116903305, + 0.0030471247155219316, + 0.005262597464025021, + 0.004743076860904694, + 0.013299054466187954, + 0.009052186273038387, + 0.017820606008172035, + 0.03577428311109543, + 0.008322648704051971, + 0.01074728462845087, + 0.02559744380414486, + 0.015568680129945278, + 0.00667258445173502, + 0.005848265253007412, + 0.008851193822920322, + 0.015153327956795692, + 0.011428432539105415, + 0.012039595283567905, + 0.052162203937768936, + 0.04302705079317093, + 0.005250377114862204, + 0.005520727019757032, + 0.008725707419216633, + 0.00954944733530283, + 0.007590932305902243, + 0.009639670141041279, + 0.00994962640106678, + 0.008512499742209911, + 0.0062140715308487415, + 0.009724907577037811, + 0.010433153249323368, + 0.009030764922499657, + 0.015641698613762856, + 0.008158775977790356, + 0.015368524007499218, + 0.00891636312007904, + 0.0069198003038764, + 0.007171230390667915, + 0.009941562078893185, + 0.00899630505591631, + 0.00792710855603218, + 0.006819205358624458, + 0.010358775034546852, + 0.008229025639593601, + 0.0103770662099123, + 0.013945206999778748, + 0.017160357907414436, + 0.02297104336321354, + 0.013143224641680717, + 0.012665423564612865, + 0.018928561359643936, + 0.01395436655730009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.47434425354003906 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043052323162555695 + ], + "opposite_logprob_tokens": [ + -3.1680524349212646 + ], + "logprob_reward": -0.043052323162555695, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043052323162555695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14661361277103424, + 0.11358363181352615, + 0.0464540459215641, + 0.023442130535840988, + 0.07037848979234695, + 0.13914340734481812, + 0.035815462470054626, + 0.0411188080906868, + 0.13418522477149963, + 0.24926519393920898 + ], + "credit_map": [ + 0.016629064455628395, + 0.03431360796093941, + 0.01618162728846073, + 0.01468473207205534, + 0.013651726767420769, + 0.016331080347299576, + 0.015043499879539013, + 0.019778279587626457, + 0.014668523333966732, + 0.012562602758407593, + 0.018828274682164192, + 0.012396483682096004, + 0.009005896747112274, + 0.013511260971426964, + 0.020778397098183632, + 0.011832191608846188, + 0.004471941385418177, + 0.004959579091519117, + 0.00405021570622921, + 0.012575128115713596, + 0.0024274459574371576, + 0.0028532070573419333, + 0.005854420829564333, + 0.009262106381356716, + 0.004237618762999773, + 0.002264183247461915, + 0.0031202221289277077, + 0.006467870902270079, + 0.0014760125195607543, + 0.001435096375644207, + 0.0024030953645706177, + 0.0020380301866680384, + 0.00690457783639431, + 0.004844686482101679, + 0.008648823946714401, + 0.015737051144242287, + 0.004958870820701122, + 0.006267457269132137, + 0.013850077986717224, + 0.009166945703327656, + 0.005388116929680109, + 0.0031775874085724354, + 0.007217442616820335, + 0.014356307685375214, + 0.007748227100819349, + 0.0074262311682105064, + 0.053354643285274506, + 0.04047485813498497, + 0.0032071645837277174, + 0.002564940834417939, + 0.006196625996381044, + 0.0054765804670751095, + 0.00357204582542181, + 0.005065449513494968, + 0.005687016062438488, + 0.004045636858791113, + 0.002616538666188717, + 0.0035608916077762842, + 0.005260562989860773, + 0.005796190816909075, + 0.004789954982697964, + 0.0049032242968678474, + 0.009089695289731026, + 0.005101749673485756, + 0.01042432151734829, + 0.009626159444451332, + 0.018759841099381447, + 0.026110561564564705, + 0.01563936285674572, + 0.010550721548497677, + 0.027266494929790497, + 0.015807755291461945, + 0.021229606121778488, + 0.027613526210188866, + 0.034886281937360764, + 0.05067533254623413, + 0.021445471793413162, + 0.0281837098300457, + 0.038618918508291245, + 0.026612339541316032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.043052323162555695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic shot begins, moving from left to right, capturing a wide shot of a bustling open-air market on a sun-drenched cobblestone square. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long, gentle shadows and creating a warm, inviting atmosphere. The camera first glides past a vendor's stall, where vibrant powdered spices—fiery red chili powder, golden turmeric, and earthy brown cumin—are piled high in large wooden bowls. Next to them, stacks of meticulously handwoven baskets made of natural straw and spandex display intricate patterns. The camera continues in a steady motion, showcasing a family gathered together; a mother, her hair tied in a loose bun and wearing a simple white linen dress, gently places a small, shimmering candy into her daughter's waiting hands. As the child savors it, her eyes widen with delight. In the background, other shoppers stroll between stalls shaded by colorful fabric awnings, their muffled conversation blending with the gentle, melodic notes of a lute played by a musician seated on a low stool. The pot finally fell on a little boy in a bright blue T-shirt, who excitedly tugged at his father's sleeve, stretched out his arm, and pointed with pure wonder at a stall overflowing with colorful, sugar-dusted candies, their surfaces gleaming in the warm light.", + "seed": 17977, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_3263_seed_17977_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": -0.23040315546095372, + "vlm_logprob": -0.23040315546095372, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.23040315546095372, + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0576090812683105 + ], + "opposite_logprob_tokens": [ + -0.43260908126831055 + ], + "logprob_reward": -1.0576090812683105, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0576090812683105, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15579088032245636, + 0.03623335808515549, + 0.06686902046203613, + 0.07520925253629684, + 0.044935788959264755, + 0.19952702522277832, + 0.07049757987260818, + 0.07361564040184021, + 0.22474482655525208, + 0.052576690912246704 + ], + "credit_map": [ + 0.013210361823439598, + 0.012388712726533413, + 0.011653459630906582, + 0.017029304057359695, + 0.009571489877998829, + 0.023092789575457573, + 0.02564726024866104, + 0.04319750890135765, + 0.0027673954609781504, + 0.004641006235033274, + 0.002699882024899125, + 0.005394486244767904, + 0.006414247676730156, + 0.005104775074869394, + 0.005278875585645437, + 0.003932689316570759, + 0.005963033996522427, + 0.008354107849299908, + 0.006637626327574253, + 0.009687266312539577, + 0.010661565698683262, + 0.007919787429273129, + 0.009255973622202873, + 0.008389662019908428, + 0.008486321195960045, + 0.008427904918789864, + 0.01201168168336153, + 0.012777227908372879, + 0.006329139694571495, + 0.014598952606320381, + 0.007298439275473356, + 0.005279580596834421, + 0.0046017966233193874, + 0.005733917001634836, + 0.0069441040977835655, + 0.008247493766248226, + 0.006066829431802034, + 0.005946126766502857, + 0.004594746977090836, + 0.0028007770888507366, + 0.01373897772282362, + 0.0198865607380867, + 0.03180520981550217, + 0.06507398188114166, + 0.020141806453466415, + 0.01466579083353281, + 0.019242258742451668, + 0.014972439967095852, + 0.008191966451704502, + 0.010645804926753044, + 0.013375254347920418, + 0.00974034983664751, + 0.01026309933513403, + 0.004618443548679352, + 0.0049503943882882595, + 0.008712269365787506, + 0.005927711259573698, + 0.012384708039462566, + 0.009461055509746075, + 0.008084302768111229, + 0.011117251589894295, + 0.00802812073379755, + 0.01072568167001009, + 0.007886810228228569, + 0.014671193435788155, + 0.01801094226539135, + 0.023407796397805214, + 0.017521658912301064, + 0.0357273668050766, + 0.024178238585591316, + 0.04443718120455742, + 0.04679044336080551, + 0.004796289838850498, + 0.005549862515181303, + 0.0068872300907969475, + 0.006138463970273733, + 0.007993899285793304, + 0.004679384175688028, + 0.009853838942945004, + 0.006677720695734024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -1.0576090812683105 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048697132617235184 + ], + "opposite_logprob_tokens": [ + -3.0486972332000732 + ], + "logprob_reward": -0.048697132617235184, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048697132617235184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1856600046157837, + 0.10187667608261108, + 0.0791483074426651, + 0.06202811375260353, + 0.09382519125938416, + 0.06421229243278503, + 0.07955999672412872, + 0.07411523163318634, + 0.19631798565387726, + 0.06325618922710419 + ], + "credit_map": [ + 0.033580996096134186, + 0.021131806075572968, + 0.021340861916542053, + 0.021810276433825493, + 0.01120627112686634, + 0.015982449054718018, + 0.021601812914013863, + 0.03900555148720741, + 0.017695195972919464, + 0.008414462208747864, + 0.016624143347144127, + 0.018077746033668518, + 0.010341807268559933, + 0.00801408477127552, + 0.010248823091387749, + 0.012460418045520782, + 0.007810050621628761, + 0.010884096845984459, + 0.008244013413786888, + 0.010751241818070412, + 0.012540078721940517, + 0.011063290759921074, + 0.009134599938988686, + 0.00872094091027975, + 0.012094075791537762, + 0.007927562110126019, + 0.0063780550844967365, + 0.006563124246895313, + 0.004998232703655958, + 0.01399967446923256, + 0.004520593676716089, + 0.005546802654862404, + 0.006616599857807159, + 0.008041433058679104, + 0.010044066235423088, + 0.00894419476389885, + 0.012007486075162888, + 0.023809870705008507, + 0.012504378333687782, + 0.011857174336910248, + 0.004137038718909025, + 0.008475868962705135, + 0.008810956962406635, + 0.011898942291736603, + 0.007381495088338852, + 0.0063265287317335606, + 0.008377076126635075, + 0.008804388344287872, + 0.00774336839094758, + 0.01031703595072031, + 0.012483510188758373, + 0.01456440519541502, + 0.009607535786926746, + 0.005493461154401302, + 0.010437363758683205, + 0.008913322351872921, + 0.006433556787669659, + 0.006851943209767342, + 0.009302002377808094, + 0.009777399711310863, + 0.01175399124622345, + 0.01157971192151308, + 0.009572161361575127, + 0.008844476193189621, + 0.01344800740480423, + 0.015575280413031578, + 0.027088738977909088, + 0.015332468785345554, + 0.026028010994195938, + 0.02063722535967827, + 0.03711448609828949, + 0.041093774139881134, + 0.006343949120491743, + 0.007643981371074915, + 0.007511114701628685, + 0.008884892798960209, + 0.007819601334631443, + 0.005858642049133778, + 0.011455358006060123, + 0.0077386596240103245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.048697132617235184 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263044033199549 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.005263044033199549, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263044033199549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21560506522655487, + 0.10044784098863602, + 0.10587123036384583, + 0.11845210939645767, + 0.06163254380226135, + 0.09698840975761414, + 0.07256940752267838, + 0.11062152683734894, + 0.053197022527456284, + 0.06461487710475922 + ], + "credit_map": [ + 0.021299852058291435, + 0.015504090115427971, + 0.02100227028131485, + 0.022875739261507988, + 0.018988585099577904, + 0.034566812217235565, + 0.035297293215990067, + 0.04607041925191879, + 0.011234529316425323, + 0.010544710792601109, + 0.012705039232969284, + 0.015120460651814938, + 0.014343074522912502, + 0.013665330596268177, + 0.010595372878015041, + 0.012239323928952217, + 0.008907217532396317, + 0.010724497027695179, + 0.016729433089494705, + 0.013494039885699749, + 0.012240144424140453, + 0.017588762566447258, + 0.013952741399407387, + 0.012234396301209927, + 0.012832148000597954, + 0.012088890187442303, + 0.019842207431793213, + 0.01975429803133011, + 0.01198534481227398, + 0.022794540971517563, + 0.011366651393473148, + 0.007788024842739105, + 0.006164827849715948, + 0.0065019819885492325, + 0.010389069095253944, + 0.012071222998201847, + 0.005716615356504917, + 0.008599267341196537, + 0.0076948958449065685, + 0.004494662396609783, + 0.008162920363247395, + 0.015287235379219055, + 0.01329550240188837, + 0.021575339138507843, + 0.010456070303916931, + 0.0074542793445289135, + 0.009065533056855202, + 0.011691523715853691, + 0.007896837778389454, + 0.013096235692501068, + 0.01092775259166956, + 0.009357644245028496, + 0.008337629027664661, + 0.004873594269156456, + 0.010413454845547676, + 0.007666260469704866, + 0.008262270130217075, + 0.013735772110521793, + 0.01686847396194935, + 0.0162282045930624, + 0.01189685519784689, + 0.010349695570766926, + 0.018659275025129318, + 0.014620988629758358, + 0.007023437414318323, + 0.005924362689256668, + 0.007519464008510113, + 0.00555826909840107, + 0.006837267894297838, + 0.007959415204823017, + 0.005176571197807789, + 0.007198232226073742, + 0.006697848904877901, + 0.004793877247720957, + 0.008941159583628178, + 0.006358222104609013, + 0.006091316696256399, + 0.01195937767624855, + 0.011445286683738232, + 0.008327788673341274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.005263044033199549 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371411934494972 + ], + "opposite_logprob_tokens": [ + -3.4087140560150146 + ], + "logprob_reward": -0.03371411934494972, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371411934494972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07732266187667847, + 0.04676381126046181, + 0.05037640035152435, + 0.04773201048374176, + 0.04732696712017059, + 0.051259227097034454, + 0.049122441560029984, + 0.16957785189151764, + 0.2989523708820343, + 0.16156630218029022 + ], + "credit_map": [ + 0.00937977060675621, + 0.010936380364000797, + 0.010138265788555145, + 0.010554627515375614, + 0.005182548891752958, + 0.006251147016882896, + 0.009364347904920578, + 0.015515577048063278, + 0.004619802813977003, + 0.00454503670334816, + 0.006245396565645933, + 0.007432796526700258, + 0.007050859276205301, + 0.007236490026116371, + 0.004733641166239977, + 0.004899783991277218, + 0.004975027404725552, + 0.006521711591631174, + 0.006628326140344143, + 0.006028709467500448, + 0.0070665208622813225, + 0.00800965167582035, + 0.005681039299815893, + 0.005465412046760321, + 0.007490375079214573, + 0.0041394466534256935, + 0.005917626433074474, + 0.0039594718255102634, + 0.0045269932597875595, + 0.013990610837936401, + 0.003642885247245431, + 0.004064599983394146, + 0.003910541068762541, + 0.004776292014867067, + 0.0053579071536660194, + 0.005348120350390673, + 0.005492138676345348, + 0.008789689280092716, + 0.006707970052957535, + 0.006944308523088694, + 0.004691843409091234, + 0.006114664021879435, + 0.006060366053134203, + 0.008891659788787365, + 0.005825412925332785, + 0.006861171685159206, + 0.007380767725408077, + 0.005433341488242149, + 0.004634689539670944, + 0.004085560794919729, + 0.00886749941855669, + 0.007134824059903622, + 0.005852896254509687, + 0.003152447286993265, + 0.009653057903051376, + 0.005741470027714968, + 0.011756302788853645, + 0.01818501017987728, + 0.0191038828343153, + 0.029703279957175255, + 0.020464729517698288, + 0.023266565054655075, + 0.028575899079442024, + 0.01852218061685562, + 0.018204420804977417, + 0.023700132966041565, + 0.03497911989688873, + 0.02294696681201458, + 0.04483745992183685, + 0.029424749314785004, + 0.06048581749200821, + 0.06437370926141739, + 0.012427558191120625, + 0.015409085899591446, + 0.014283274300396442, + 0.01651078090071678, + 0.03349791094660759, + 0.015409838408231735, + 0.0346490740776062, + 0.019378777593374252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.03371411934494972 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16836048662662506, + 0.0709514245390892, + 0.06036670133471489, + 0.041125282645225525, + 0.07286135107278824, + 0.0785607323050499, + 0.07825570553541183, + 0.12810780107975006, + 0.11244216561317444, + 0.18896833062171936 + ], + "credit_map": [ + 0.012117884121835232, + 0.016817526891827583, + 0.014407200738787651, + 0.02439277060329914, + 0.01079857163131237, + 0.02365686744451523, + 0.024264369159936905, + 0.04190529137849808, + 0.00931487139314413, + 0.006317722611129284, + 0.009898676536977291, + 0.011259160935878754, + 0.008912788704037666, + 0.008030048571527004, + 0.007971110753715038, + 0.009247043170034885, + 0.004781270399689674, + 0.007511637639254332, + 0.007714021485298872, + 0.009089435450732708, + 0.009562553837895393, + 0.008294889703392982, + 0.00607988191768527, + 0.007333014160394669, + 0.0033447486348450184, + 0.0050026727840304375, + 0.008555206470191479, + 0.00551079073920846, + 0.004086263012140989, + 0.0038608333561569452, + 0.005296286195516586, + 0.005468482617288828, + 0.005932117346674204, + 0.008305889554321766, + 0.009423231706023216, + 0.010925239883363247, + 0.007761345710605383, + 0.014149222522974014, + 0.00880517903715372, + 0.0075591248460114, + 0.005167094059288502, + 0.012259413488209248, + 0.010421006008982658, + 0.022008350118994713, + 0.009454785846173763, + 0.005464726127684116, + 0.008708038367331028, + 0.005077318754047155, + 0.007164042443037033, + 0.012526336126029491, + 0.01174730435013771, + 0.01185088325291872, + 0.008473142981529236, + 0.006438193842768669, + 0.011654718779027462, + 0.008401080034673214, + 0.011249277740716934, + 0.013024256564676762, + 0.018857399001717567, + 0.020847463980317116, + 0.016333051025867462, + 0.014721838757395744, + 0.019827252253890038, + 0.01324725802987814, + 0.011817163787782192, + 0.011379011906683445, + 0.014691601507365704, + 0.012521814554929733, + 0.018633287400007248, + 0.013523629866540432, + 0.013014205731451511, + 0.01686144433915615, + 0.023750731721520424, + 0.01894332654774189, + 0.024557974189519882, + 0.023122254759073257, + 0.022790243849158287, + 0.018635420128703117, + 0.026053592562675476, + 0.031114792451262474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006732400041073561 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic shot begins, moving from left to right, capturing a wide shot of a bustling open-air market on a sun-drenched cobblestone square. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long, gentle shadows and creating a warm, inviting atmosphere. The camera first glides past a vendor's stall, where vibrant powdered spices—fiery red chili powder, golden turmeric, and earthy brown cumin—are piled high in large wooden bowls. Next to them, stacks of meticulously handwoven baskets made of natural straw and spandex display intricate patterns. The camera continues in a steady motion, showcasing a family gathered together; a mother, her hair tied in a loose bun and wearing a simple white linen dress, gently places a small, shimmering candy into her daughter's waiting hands. As the child savors it, her eyes widen with delight. In the background, other shoppers stroll between stalls shaded by colorful fabric awnings, their muffled conversation blending with the gentle, melodic notes of a lute played by a musician seated on a low stool. The pot finally fell on a little boy in a bright blue T-shirt, who excitedly tugged at his father's sleeve, stretched out his arm, and pointed with pure wonder at a stall overflowing with colorful, sugar-dusted candies, their surfaces gleaming in the warm light.", + "seed": 17978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_11/index_3263_seed_17978_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": -0.09047623686492443, + "vlm_logprob": -0.09047623686492443, + "vlm_yes_no_margin": 3.399999976158142, + "vlm_token_credit": -0.09047623686492443, + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1363714635372162 + ], + "opposite_logprob_tokens": [ + -2.136371374130249 + ], + "logprob_reward": -0.1363714635372162, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1363714635372162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03076294995844364, + 0.005315229296684265, + 0.009680315852165222, + 0.007679400499910116, + 0.004781829193234444, + 0.8596354126930237, + 0.009217617101967335, + 0.03151515871286392, + 0.023664986714720726, + 0.01774713024497032 + ], + "credit_map": [ + 0.0036309142597019672, + 0.003711786586791277, + 0.005847186781466007, + 0.003232874209061265, + 0.0022919350303709507, + 0.003329599043354392, + 0.0032609179615974426, + 0.005457737948745489, + 0.0008573841769248247, + 0.0004288835043553263, + 0.0005798001657240093, + 0.0009250263101421297, + 0.0004815831780433655, + 0.0003794319636654109, + 0.000849926786031574, + 0.0008131939102895558, + 0.001513661234639585, + 0.0009323967969976366, + 0.001572879496961832, + 0.0014012539759278297, + 0.0009426430915482342, + 0.0008367456030100584, + 0.001216087257489562, + 0.0012646493269130588, + 0.0008110920898616314, + 0.0010722096776589751, + 0.0008096941164694726, + 0.0007293823873624206, + 0.001109906705096364, + 0.0007240368286147714, + 0.0017234256956726313, + 0.0006996536976657808, + 0.00037212675670161843, + 0.0005819109501317143, + 0.0006042807945050299, + 0.0008416405180469155, + 0.0008608315256424248, + 0.0005167533527128398, + 0.0004487284168135375, + 0.0005555576062761247, + 0.054274629801511765, + 0.04735657200217247, + 0.048966024070978165, + 0.3722008466720581, + 0.019020460546016693, + 0.0305938757956028, + 0.0370076559484005, + 0.2502153813838959, + 0.0007633898640051484, + 0.0009399057598784566, + 0.0020625588949769735, + 0.00178860395681113, + 0.0006018245476298034, + 0.0004157483926974237, + 0.0010778135620057583, + 0.0015677729388698936, + 0.004229506477713585, + 0.006892934907227755, + 0.004908578936010599, + 0.0042386422865092754, + 0.0032088502775877714, + 0.0025641859974712133, + 0.002784313401207328, + 0.002688149455934763, + 0.002478683367371559, + 0.004578977357596159, + 0.0032427741680294275, + 0.003844637656584382, + 0.003034074092283845, + 0.0018755010096356273, + 0.0019943618681281805, + 0.002615978242829442, + 0.0019039073958992958, + 0.0014190967194736004, + 0.0034772329963743687, + 0.0035146502777934074, + 0.0020683323964476585, + 0.001756990677677095, + 0.00149958161637187, + 0.0021073385141789913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.1363714635372162 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498848482966423 + ], + "opposite_logprob_tokens": [ + -2.929988384246826 + ], + "logprob_reward": -0.05498848482966423, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498848482966423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021861359477043152, + 0.014790374785661697, + 0.026476558297872543, + 0.014459118247032166, + 0.005875210743397474, + 0.892304539680481, + 0.006723498925566673, + 0.0030364671256393194, + 0.006090077571570873, + 0.008382759988307953 + ], + "credit_map": [ + 0.002226036274805665, + 0.0029087995644658804, + 0.006232275627553463, + 0.001444010529667139, + 0.001510041649453342, + 0.001504329266026616, + 0.0021978244185447693, + 0.0038380404002964497, + 0.0018135170685127378, + 0.00143460463732481, + 0.0021740784868597984, + 0.0035885528195649385, + 0.0009017933625727892, + 0.0007825115462765098, + 0.001985553652048111, + 0.0021097627468407154, + 0.0013762289891019464, + 0.0008942094864323735, + 0.010161885991692543, + 0.009574875235557556, + 0.0007550914888270199, + 0.0008129210327751935, + 0.0014292087871581316, + 0.0014721372863277793, + 0.0014584118034690619, + 0.0013363425387069583, + 0.0015196851454675198, + 0.0011210995726287365, + 0.0021708097774535418, + 0.001253737136721611, + 0.004120497032999992, + 0.0014785362873226404, + 0.0005687419325113297, + 0.00076963851461187, + 0.0006061392487026751, + 0.0006795093067921698, + 0.001251010806299746, + 0.0005784374661743641, + 0.0008290328551083803, + 0.0005927005549892783, + 0.05576825141906738, + 0.04923820123076439, + 0.051427531987428665, + 0.3863496482372284, + 0.020065730437636375, + 0.03167949616909027, + 0.03843938559293747, + 0.2593362629413605, + 0.0007140252855606377, + 0.0014671401586383581, + 0.0006617679609917104, + 0.0010260295821353793, + 0.0006087836227379739, + 0.00047245551832020283, + 0.0010217074304819107, + 0.0007515894249081612, + 0.00043582121725194156, + 0.0004195363726466894, + 0.0003490434028208256, + 0.0004664267471525818, + 0.00031203479738906026, + 0.0002859294763766229, + 0.00038520386442542076, + 0.0003824713930953294, + 0.0006774214562028646, + 0.0015740329399704933, + 0.0005753637524321675, + 0.0010016842279583216, + 0.0006004511378705502, + 0.0003671989543363452, + 0.0005871669272892177, + 0.0007067580590955913, + 0.0007402013870887458, + 0.001421917462721467, + 0.0014457358047366142, + 0.001765242894180119, + 0.0006232726154848933, + 0.00047166176955215633, + 0.0007851988193579018, + 0.0011295299045741558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.05498848482966423 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261858459562063 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005261858459562063, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261858459562063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06999707221984863, + 0.05476992577314377, + 0.08131357282400131, + 0.0203329436480999, + 0.026759298518300056, + 0.6079741716384888, + 0.0309460386633873, + 0.06290356069803238, + 0.028416043147444725, + 0.016587385907769203 + ], + "credit_map": [ + 0.007262916304171085, + 0.013132522813975811, + 0.012677011080086231, + 0.00891427043825388, + 0.0037053159903734922, + 0.004878105595707893, + 0.00807597953826189, + 0.011350945569574833, + 0.005210466217249632, + 0.004885729402303696, + 0.004659351892769337, + 0.012908028438687325, + 0.007003705482929945, + 0.004217916168272495, + 0.006748740561306477, + 0.009135986678302288, + 0.006133762653917074, + 0.00527875917032361, + 0.02811860665678978, + 0.02701599709689617, + 0.0023332543205469847, + 0.0026744194328784943, + 0.004522989504039288, + 0.005235779099166393, + 0.0019311178475618362, + 0.0016251055058091879, + 0.0026391036808490753, + 0.0026712906546890736, + 0.0025666847359389067, + 0.0014449878362938762, + 0.004991885740309954, + 0.0024627672974020243, + 0.003279609140008688, + 0.002942199818789959, + 0.006117335986346006, + 0.0042815194465219975, + 0.0023756795562803745, + 0.002140380209311843, + 0.003146623494103551, + 0.0024759487714618444, + 0.03676128014922142, + 0.03364114835858345, + 0.035102471709251404, + 0.2642556428909302, + 0.013144559226930141, + 0.021404188126325607, + 0.026954537257552147, + 0.1767103523015976, + 0.0025684740394353867, + 0.0017903195694088936, + 0.010023808106780052, + 0.002585054375231266, + 0.001953636994585395, + 0.0016509752022102475, + 0.0020780886989086866, + 0.008295678533613682, + 0.016918279230594635, + 0.0105446120724082, + 0.008130998350679874, + 0.006635130383074284, + 0.00460763880982995, + 0.0034521964844316244, + 0.006164653692394495, + 0.006450054235756397, + 0.004262607544660568, + 0.004333313554525375, + 0.003476360347121954, + 0.006003755144774914, + 0.002767880680039525, + 0.002491766819730401, + 0.0021240597125142813, + 0.0029562993440777063, + 0.0024577099829912186, + 0.0015671848086640239, + 0.00242150598205626, + 0.003446141956374049, + 0.0019530856516212225, + 0.001648346078582108, + 0.0011298446916043758, + 0.0019635672215372324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.005261858459562063 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521456480026245 + ], + "opposite_logprob_tokens": [ + -1.5021456480026245 + ], + "logprob_reward": -0.2521456480026245, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521456480026245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07012532651424408, + 0.06088420748710632, + 0.1455772966146469, + 0.04208395630121231, + 0.07088221609592438, + 0.38509345054626465, + 0.03443559631705284, + 0.05147429183125496, + 0.11188492178916931, + 0.027558857575058937 + ], + "credit_map": [ + 0.008746486157178879, + 0.011776660569012165, + 0.010225661098957062, + 0.010254884138703346, + 0.005625439342111349, + 0.005374915432184935, + 0.007684681098908186, + 0.0104365898296237, + 0.010426703840494156, + 0.004380383063107729, + 0.004175300244241953, + 0.012212034314870834, + 0.004122378770262003, + 0.0032337976153939962, + 0.011319508776068687, + 0.011014089919626713, + 0.007230413146317005, + 0.006218799855560064, + 0.059728898108005524, + 0.05728389322757721, + 0.002102537779137492, + 0.003079629736021161, + 0.006116872653365135, + 0.0038162388373166323, + 0.0031866957433521748, + 0.004655846860259771, + 0.005266000051051378, + 0.005518870893865824, + 0.007819059304893017, + 0.004445397295057774, + 0.006900385953485966, + 0.004291694611310959, + 0.004904631525278091, + 0.009141180664300919, + 0.007197361905127764, + 0.013087859377264977, + 0.015437306836247444, + 0.0076213194988667965, + 0.006904930807650089, + 0.006587620824575424, + 0.023249445483088493, + 0.02077382616698742, + 0.02187161333858967, + 0.16475088894367218, + 0.009971480816602707, + 0.014493104070425034, + 0.018474923446774483, + 0.11150810122489929, + 0.0024114223197102547, + 0.007027294486761093, + 0.005830667447298765, + 0.00481007993221283, + 0.002961312886327505, + 0.0022009483072906733, + 0.003966072574257851, + 0.005227796267718077, + 0.009968902915716171, + 0.009939377196133137, + 0.006245127413421869, + 0.00810485053807497, + 0.004401919431984425, + 0.0028473890852183104, + 0.005190498195588589, + 0.00477622402831912, + 0.011969161219894886, + 0.029436636716127396, + 0.010812673717737198, + 0.016049053519964218, + 0.011600389145314693, + 0.005299696233123541, + 0.01447659358382225, + 0.012240709736943245, + 0.002103137318044901, + 0.0036053035873919725, + 0.0046416777186095715, + 0.0061133988201618195, + 0.002884235233068466, + 0.0018537766300141811, + 0.002877765102311969, + 0.0034795603714883327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.2521456480026245 + }, + { + "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 bustling open-air market on a sun-drenched cobblestone square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move in a steady motion from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant powdered spices (red chili powder, golden turmeric, and brown cumin) visible in large wooden bowls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a family shown where a mother in a white linen dress gives a candy to her delighted daughter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a little boy in a bright blue T-shirt shown pointing at a stall overflowing with colorful, sugar-dusted candies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036137294955551624 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036137294955551624, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036137294955551624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013203756883740425, + 0.013824662193655968, + 0.024604853242635727, + 0.007690471597015858, + 0.013678385876119137, + 0.840907633304596, + 0.03002850152552128, + 0.021439678966999054, + 0.024220362305641174, + 0.010401601903140545 + ], + "credit_map": [ + 0.0018844801234081388, + 0.0016731603536754847, + 0.0015502253081649542, + 0.0018388285534456372, + 0.0012170287081971765, + 0.0013550863368436694, + 0.0019537489861249924, + 0.0017311997944489121, + 0.0009614875889383256, + 0.0012398525141179562, + 0.0013112211599946022, + 0.0039002359844744205, + 0.0010317815467715263, + 0.0009137587039731443, + 0.0019391763489693403, + 0.002527149161323905, + 0.0021278231870383024, + 0.0018790275789797306, + 0.008241270668804646, + 0.0077117509208619595, + 0.0010437669698148966, + 0.0009750723838806152, + 0.0013656762894243002, + 0.001260467222891748, + 0.0008483656565658748, + 0.0007904042140580714, + 0.0009685634286142886, + 0.001047235680744052, + 0.001748481416143477, + 0.0006073398981243372, + 0.001097851200029254, + 0.000582230684813112, + 0.001238940516486764, + 0.0015829455805942416, + 0.002485310658812523, + 0.002784005831927061, + 0.001444819732569158, + 0.0014739412581548095, + 0.0010800338350236416, + 0.001588391256518662, + 0.05254434794187546, + 0.04705803468823433, + 0.048843201249837875, + 0.36381393671035767, + 0.01913956180214882, + 0.029640264809131622, + 0.03616148233413696, + 0.24370692670345306, + 0.0023012314923107624, + 0.002324985805898905, + 0.009116223081946373, + 0.003127738833427429, + 0.0016398977022618055, + 0.0013148004654794931, + 0.0019546940457075834, + 0.008248933590948582, + 0.0030293664894998074, + 0.005557445343583822, + 0.0024555912241339684, + 0.0028287204913794994, + 0.002115016570314765, + 0.0012335615465417504, + 0.002278233878314495, + 0.0019417466828599572, + 0.0037211556918919086, + 0.006405536085367203, + 0.002824379364028573, + 0.003120812587440014, + 0.002106576692312956, + 0.0010529879946261644, + 0.002700529294088483, + 0.0022883876226842403, + 0.0016035110456869006, + 0.0014921765541657805, + 0.000798481865786016, + 0.0017480339156463742, + 0.0011237967992201447, + 0.000989135936833918, + 0.001410117605701089, + 0.001236348762176931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.0036137294955551624 + } + ], + "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/0000017/group_12/index_tensor([2089, 2089])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_tensor([2089, 2089])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..26daf255940cee9f64806679e7f20e0de1fcb1ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_tensor([2089, 2089])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a bird's-eye view, a static extreme close-up captures a delicate, shimmering film of soapy water spreading across the smooth, dark surface of a polished granite countertop. Under soft, diffuse studio light, the film's surface is a captivating swirl of iridescent pink, cyan, and gold that slowly shifts and flows. The film thins noticeably, the colors swirling at an increasingly rapid pace until a breaking point appears near the center. Instantly, the fragile membrane fails, catastrophically collapsing inward from all sides as surface tension breaks down. The liquid coalesces, rushing swiftly and decisively toward the center in a chaotic surge that rapidly organizes and pulls the water into a nearly perfectly circular shape. The movement ceases almost as quickly as it began, leaving a clean, transparent wet ring on the countertop's smooth surface, slightly thicker at the edges. A few tiny, lingering bubbles shimmer within the clear liquid, creating a serene and mesmerizing vision of microscopic physics.", + "seed": 9953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_2089_seed_9953_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": -0.05619501060573384, + "vlm_logprob": -0.05619501060573384, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.05619501060573384, + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11095915734767914 + ], + "opposite_logprob_tokens": [ + -2.360959053039551 + ], + "logprob_reward": -0.11095915734767914, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11095915734767914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06861591339111328, + 0.05065315216779709, + 0.03864772617816925, + 0.0528729073703289, + 0.06760602444410324, + 0.05396106839179993, + 0.04737528786063194, + 0.053582608699798584, + 0.14022010564804077, + 0.4264652729034424 + ], + "credit_map": [ + 0.005627287086099386, + 0.008576272055506706, + 0.006261355709284544, + 0.009521732106804848, + 0.009030786342918873, + 0.01001169253140688, + 0.007744066417217255, + 0.011842717416584492, + 0.005730590783059597, + 0.006209016777575016, + 0.006669250782579184, + 0.007926723919808865, + 0.007273402065038681, + 0.004224490839987993, + 0.0070005920715630054, + 0.0056190793402493, + 0.003181896172463894, + 0.005652218591421843, + 0.002838578075170517, + 0.00560890743508935, + 0.007439330220222473, + 0.0043482850305736065, + 0.005769113078713417, + 0.0038093971088528633, + 0.006630914285778999, + 0.008728285320103168, + 0.005805560387670994, + 0.004770332481712103, + 0.005207487847656012, + 0.00648978166282177, + 0.006717531476169825, + 0.008523005992174149, + 0.006551278289407492, + 0.008264713920652866, + 0.008734417147934437, + 0.007518583443015814, + 0.012143982574343681, + 0.007222308777272701, + 0.009897295385599136, + 0.007273443043231964, + 0.007585349027067423, + 0.005346044432371855, + 0.009248077869415283, + 0.004922329448163509, + 0.006017098668962717, + 0.005191405303776264, + 0.009813536889851093, + 0.005837224889546633, + 0.005813145078718662, + 0.005277062766253948, + 0.00621793232858181, + 0.008494140580296516, + 0.004990267101675272, + 0.0037883948534727097, + 0.00809460412710905, + 0.004699739161878824, + 0.008689620532095432, + 0.005414375104010105, + 0.012630628421902657, + 0.005745657254010439, + 0.005652532447129488, + 0.005067575257271528, + 0.0051255542784929276, + 0.005256659351289272, + 0.01681571640074253, + 0.024520058184862137, + 0.01634606532752514, + 0.01112730335444212, + 0.03585422784090042, + 0.010463515296578407, + 0.013854239135980606, + 0.011238973587751389, + 0.07828280329704285, + 0.04794665798544884, + 0.10704808682203293, + 0.02257891371846199, + 0.10912860929965973, + 0.015703631564974785, + 0.03002917394042015, + 0.015747351571917534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.11095915734767914 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036444902420044 + ], + "opposite_logprob_tokens": [ + -1.9103643894195557 + ], + "logprob_reward": -0.16036444902420044, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036444902420044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650243043899536, + 0.06173583120107651, + 0.05548710376024246, + 0.056143540889024734, + 0.04065535590052605, + 0.06866933405399323, + 0.03607085347175598, + 0.054039135575294495, + 0.19311800599098206, + 0.16905653476715088 + ], + "credit_map": [ + 0.017813745886087418, + 0.025103658437728882, + 0.027008237317204475, + 0.055358242243528366, + 0.0297640860080719, + 0.033935289829969406, + 0.033369701355695724, + 0.04267134144902229, + 0.004839605651795864, + 0.010502098128199577, + 0.006389017216861248, + 0.008212991990149021, + 0.013416986912488937, + 0.003283627098426223, + 0.009732332080602646, + 0.005359170492738485, + 0.005492410622537136, + 0.007089747581630945, + 0.003626230638474226, + 0.0094712870195508, + 0.01137498114258051, + 0.005009710323065519, + 0.009992816485464573, + 0.003429918549954891, + 0.006466354243457317, + 0.009798616170883179, + 0.009078900329768658, + 0.006172787863761187, + 0.004312880802899599, + 0.006575302686542273, + 0.005888352170586586, + 0.007850348018109798, + 0.005755196791142225, + 0.005133196245878935, + 0.005028942134231329, + 0.0037869554944336414, + 0.007103762123733759, + 0.0037218136712908745, + 0.004029089584946632, + 0.006096400320529938, + 0.010203317739069462, + 0.009162765927612782, + 0.010367752984166145, + 0.008366450667381287, + 0.006118733901530504, + 0.005935351829975843, + 0.009811976924538612, + 0.008702992461621761, + 0.0037261582911014557, + 0.004501524847000837, + 0.0049804700538516045, + 0.005907931365072727, + 0.0035257223062217236, + 0.0030831447802484035, + 0.006280963774770498, + 0.004064940847456455, + 0.006482607219368219, + 0.006041220389306545, + 0.007255157921463251, + 0.008487793616950512, + 0.005500501021742821, + 0.007181552238762379, + 0.00669098598882556, + 0.006399317644536495, + 0.02405146136879921, + 0.032376278191804886, + 0.021187763661146164, + 0.013795105740427971, + 0.05369715765118599, + 0.012931106612086296, + 0.01958492025732994, + 0.015494213439524174, + 0.031959086656570435, + 0.018549395725131035, + 0.042815931141376495, + 0.009222553111612797, + 0.04101218655705452, + 0.0067292870953679085, + 0.0109741585329175, + 0.007793945726007223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.16036444902420044 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007257211836986244 + ], + "opposite_logprob_tokens": [ + -7.250725746154785 + ], + "logprob_reward": -0.0007257211836986244, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007257211836986244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21236465871334076, + 0.05275929719209671, + 0.04149654880166054, + 0.06554152816534042, + 0.07999802380800247, + 0.03545984625816345, + 0.02911289408802986, + 0.026520345360040665, + 0.23522605001926422, + 0.2215208113193512 + ], + "credit_map": [ + 0.021160831674933434, + 0.02028878591954708, + 0.024325331673026085, + 0.022304026409983635, + 0.03742213174700737, + 0.02385127730667591, + 0.027060260996222496, + 0.03595201298594475, + 0.004385614302009344, + 0.007991311140358448, + 0.006050029303878546, + 0.006124458741396666, + 0.012930868193507195, + 0.0037496488075703382, + 0.00766898225992918, + 0.0038583811838179827, + 0.003994072787463665, + 0.004949421156197786, + 0.0031634026672691107, + 0.00782569870352745, + 0.007117822300642729, + 0.0042112902738153934, + 0.005440892651677132, + 0.004793946165591478, + 0.0065949587151408195, + 0.010066349059343338, + 0.007862074300646782, + 0.007394417189061642, + 0.005441179033368826, + 0.008261446841061115, + 0.009906359016895294, + 0.010014744475483894, + 0.007947850972414017, + 0.013186123222112656, + 0.010944246314466, + 0.0071363565512001514, + 0.015000652521848679, + 0.008786695078015327, + 0.008930735290050507, + 0.008065366186201572, + 0.005459651816636324, + 0.004780037794262171, + 0.004866927862167358, + 0.004907389171421528, + 0.003199714468792081, + 0.003077947534620762, + 0.004230146296322346, + 0.004938032012432814, + 0.002752136904746294, + 0.0032895493786782026, + 0.0028435790445655584, + 0.003945270087569952, + 0.0027676622848957777, + 0.005479992367327213, + 0.004306945018470287, + 0.0037277585361152887, + 0.003595649963244796, + 0.003664028597995639, + 0.004427571780979633, + 0.0030838262755423784, + 0.002874196507036686, + 0.002411304507404566, + 0.003712634090334177, + 0.002751133404672146, + 0.0293811596930027, + 0.03909372165799141, + 0.026647118851542473, + 0.016291478648781776, + 0.06802912801504135, + 0.015177950263023376, + 0.023074092343449593, + 0.017531393095850945, + 0.042629070580005646, + 0.023146625608205795, + 0.05768276005983353, + 0.011338788084685802, + 0.05222728103399277, + 0.008785433135926723, + 0.015001460909843445, + 0.01070940587669611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0007257211836986244 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14789623022079468, + 0.055182624608278275, + 0.08712771534919739, + 0.034774016588926315, + 0.11086779832839966, + 0.09078547358512878, + 0.048357658088207245, + 0.04394655302166939, + 0.19966666400432587, + 0.1813952475786209 + ], + "credit_map": [ + 0.01263194065541029, + 0.015636878088116646, + 0.016104068607091904, + 0.015498246997594833, + 0.02879404090344906, + 0.016951419413089752, + 0.01955455355346203, + 0.022725077345967293, + 0.005456213839352131, + 0.007554294541478157, + 0.005621702875941992, + 0.006480157375335693, + 0.010625398717820644, + 0.0033659932669252157, + 0.011427141726016998, + 0.004651721101254225, + 0.009007212705910206, + 0.010895410552620888, + 0.007636030670255423, + 0.017282996326684952, + 0.015015571378171444, + 0.007073987275362015, + 0.012647580355405807, + 0.007568927947431803, + 0.0033043709117919207, + 0.006561786402016878, + 0.0036184722557663918, + 0.003706220770254731, + 0.00455758860334754, + 0.005088720936328173, + 0.004083307925611734, + 0.0038535494823008776, + 0.011147000826895237, + 0.017203953117132187, + 0.013939278200268745, + 0.010051414370536804, + 0.0194482933729887, + 0.0111721595749259, + 0.012900633737444878, + 0.01500506978482008, + 0.008177293464541435, + 0.01321267057210207, + 0.008214273490011692, + 0.01009330153465271, + 0.009652074426412582, + 0.006496230605989695, + 0.0249184537678957, + 0.010021179914474487, + 0.004859179258346558, + 0.006646604277193546, + 0.004535998683422804, + 0.005405751056969166, + 0.007230870425701141, + 0.004876566585153341, + 0.009768026880919933, + 0.0050346581265330315, + 0.005553916562348604, + 0.006611386779695749, + 0.005954300053417683, + 0.004851512145251036, + 0.006018342450261116, + 0.004036972764879465, + 0.005763478111475706, + 0.0051566422916948795, + 0.02460889331996441, + 0.03531251475214958, + 0.02164332941174507, + 0.014894641935825348, + 0.05140341445803642, + 0.014004126191139221, + 0.020427634939551353, + 0.017372125759720802, + 0.03539038822054863, + 0.018673062324523926, + 0.04712550342082977, + 0.009430750273168087, + 0.04100721701979637, + 0.007169141434133053, + 0.01182885468006134, + 0.010770339518785477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "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 the camera angle a static bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006728729233145714 + ], + "opposite_logprob_tokens": [ + -5.006728649139404 + ], + "logprob_reward": -0.006728729233145714, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006728729233145714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030906917527318, + 0.027754327282309532, + 0.014779296703636646, + 0.015119235031306744, + 0.0361967608332634, + 0.06904381513595581, + 0.028528062626719475, + 0.01896519400179386, + 0.24864283204078674, + 0.5100635886192322 + ], + "credit_map": [ + 0.0036861936096102, + 0.003398942295461893, + 0.00281149847432971, + 0.004859669134020805, + 0.005791984498500824, + 0.0034070154651999474, + 0.004375810734927654, + 0.0025758049450814724, + 0.0020495238713920116, + 0.004030962940305471, + 0.0030143074691295624, + 0.003337898524478078, + 0.006245389115065336, + 0.001621459610760212, + 0.004969962872564793, + 0.002484821481630206, + 0.0014411975862458348, + 0.0016183300176635385, + 0.0012582858325913548, + 0.0029340474866330624, + 0.0025112531147897243, + 0.0013690975029021502, + 0.002177025657147169, + 0.0014700595056638122, + 0.001703449059277773, + 0.0021455660462379456, + 0.0024901663418859243, + 0.0018007159233093262, + 0.0015252047451213002, + 0.0016696978127583861, + 0.002010646741837263, + 0.0017737887101247907, + 0.00334542291238904, + 0.005746832583099604, + 0.005355046130716801, + 0.003965119831264019, + 0.006099415477365255, + 0.0035061733797192574, + 0.004266458563506603, + 0.0039122942835092545, + 0.0057038296945393085, + 0.008725207298994064, + 0.007200250867754221, + 0.006814366672188044, + 0.00831538438796997, + 0.0067606121301651, + 0.016891608014702797, + 0.008632555603981018, + 0.0022835799027234316, + 0.004143210593611002, + 0.0032720020972192287, + 0.0030308589339256287, + 0.0029413150623440742, + 0.004601779859513044, + 0.004317460115998983, + 0.0039378572255373, + 0.002317377831786871, + 0.003284003585577011, + 0.00203177728690207, + 0.002495327964425087, + 0.0021454691886901855, + 0.0022257016971707344, + 0.0021636025048792362, + 0.0023019358050078154, + 0.0314658097922802, + 0.041738111525774, + 0.029488885775208473, + 0.01787472330033779, + 0.07164352387189865, + 0.01582038775086403, + 0.022056221961975098, + 0.018555153161287308, + 0.09732675552368164, + 0.054302435368299484, + 0.13179676234722137, + 0.026231035590171814, + 0.12223447859287262, + 0.019337117671966553, + 0.034722037613391876, + 0.024112924933433533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.006728729233145714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a bird's-eye view, a static extreme close-up captures a delicate, shimmering film of soapy water spreading across the smooth, dark surface of a polished granite countertop. Under soft, diffuse studio light, the film's surface is a captivating swirl of iridescent pink, cyan, and gold that slowly shifts and flows. The film thins noticeably, the colors swirling at an increasingly rapid pace until a breaking point appears near the center. Instantly, the fragile membrane fails, catastrophically collapsing inward from all sides as surface tension breaks down. The liquid coalesces, rushing swiftly and decisively toward the center in a chaotic surge that rapidly organizes and pulls the water into a nearly perfectly circular shape. The movement ceases almost as quickly as it began, leaving a clean, transparent wet ring on the countertop's smooth surface, slightly thicker at the edges. A few tiny, lingering bubbles shimmer within the clear liquid, creating a serene and mesmerizing vision of microscopic physics.", + "seed": 9954, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_2089_seed_9954_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": -0.0460361682693474, + "vlm_logprob": -0.0460361682693474, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0460361682693474, + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12021724134683609 + ], + "opposite_logprob_tokens": [ + -2.2452173233032227 + ], + "logprob_reward": -0.12021724134683609, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12021724134683609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11495248228311539, + 0.1048443540930748, + 0.05762194097042084, + 0.0889851376414299, + 0.13041771948337555, + 0.1634351909160614, + 0.15269382297992706, + 0.05834442004561424, + 0.07524570822715759, + 0.05345919355750084 + ], + "credit_map": [ + 0.011377787217497826, + 0.016446677967905998, + 0.01342698559165001, + 0.01693142205476761, + 0.013577069155871868, + 0.013183948583900928, + 0.013283733278512955, + 0.016724860295653343, + 0.02370700053870678, + 0.012903045862913132, + 0.017528878524899483, + 0.00879509188234806, + 0.011121634393930435, + 0.010459220968186855, + 0.009010788053274155, + 0.011318691074848175, + 0.006592500489205122, + 0.006589334458112717, + 0.006259486544877291, + 0.004550193902105093, + 0.00715082511305809, + 0.014411219395697117, + 0.005744537338614464, + 0.006323849316686392, + 0.010965991765260696, + 0.014887124300003052, + 0.020557951182127, + 0.00912259891629219, + 0.008785862475633621, + 0.008488672785460949, + 0.008568516932427883, + 0.0076084197498857975, + 0.013500623404979706, + 0.020789440721273422, + 0.017578383907675743, + 0.018246419727802277, + 0.013465778902173042, + 0.01684425212442875, + 0.017260722815990448, + 0.012732095085084438, + 0.0146480742841959, + 0.02058766596019268, + 0.01875326968729496, + 0.02852511964738369, + 0.021817384287714958, + 0.020721781998872757, + 0.02292918786406517, + 0.01545270811766386, + 0.016082607209682465, + 0.014520051889121532, + 0.014810551889240742, + 0.027143267914652824, + 0.01605760119855404, + 0.019800523295998573, + 0.028286896646022797, + 0.01599232107400894, + 0.006470584776252508, + 0.005176758859306574, + 0.007283270359039307, + 0.007188009098172188, + 0.0056263781152665615, + 0.008128401823341846, + 0.011255220510065556, + 0.007215794641524553, + 0.00706640537828207, + 0.010825619101524353, + 0.008061331696808338, + 0.014905769377946854, + 0.008596843108534813, + 0.007816162891685963, + 0.011158706620335579, + 0.006814872846007347, + 0.0063011739403009415, + 0.007421465590596199, + 0.005210386589169502, + 0.007274575065821409, + 0.007901289500296116, + 0.00528990663588047, + 0.008955881930887699, + 0.005104509182274342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.12021724134683609 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899297028779984 + ], + "opposite_logprob_tokens": [ + -2.578993082046509 + ], + "logprob_reward": -0.07899297028779984, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899297028779984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1474527269601822, + 0.07622619718313217, + 0.06184126436710358, + 0.0426497645676136, + 0.08565076440572739, + 0.106022909283638, + 0.13602058589458466, + 0.11761578917503357, + 0.17261089384555817, + 0.053909048438072205 + ], + "credit_map": [ + 0.011698510497808456, + 0.01604967564344406, + 0.018675707280635834, + 0.01800931990146637, + 0.017337776720523834, + 0.01783420890569687, + 0.019045935943722725, + 0.02880159765481949, + 0.006897575221955776, + 0.01302917581051588, + 0.007680258713662624, + 0.007829932495951653, + 0.011167499236762524, + 0.011431258171796799, + 0.007366249803453684, + 0.010824242606759071, + 0.005603378638625145, + 0.00803715456277132, + 0.007329516112804413, + 0.009248589165508747, + 0.009337464347481728, + 0.00990524422377348, + 0.00633525475859642, + 0.006044662091881037, + 0.005885587073862553, + 0.005868731066584587, + 0.005790220573544502, + 0.006223380099982023, + 0.003996711689978838, + 0.005425179377198219, + 0.004816863685846329, + 0.004643090069293976, + 0.007795810233801603, + 0.013881836086511612, + 0.012183158658444881, + 0.011469573713839054, + 0.009404117241501808, + 0.011898628436028957, + 0.010146764107048512, + 0.008870880119502544, + 0.013159172609448433, + 0.014101678505539894, + 0.009613445959985256, + 0.01624433323740959, + 0.01145845465362072, + 0.013620729558169842, + 0.01405251957476139, + 0.013772574253380299, + 0.01382703147828579, + 0.01859600469470024, + 0.014397148974239826, + 0.023471089079976082, + 0.010410705581307411, + 0.018896780908107758, + 0.023382291197776794, + 0.01303953304886818, + 0.012525695376098156, + 0.01151407789438963, + 0.011624329723417759, + 0.020969102159142494, + 0.015159894712269306, + 0.015509910881519318, + 0.018195217475295067, + 0.012117571197450161, + 0.018834033980965614, + 0.017041249200701714, + 0.015442834235727787, + 0.021134519949555397, + 0.024060647934675217, + 0.02246413379907608, + 0.03663118928670883, + 0.017002297565340996, + 0.006251585204154253, + 0.005156521685421467, + 0.005188716575503349, + 0.006933713797479868, + 0.006625031586736441, + 0.005974313709884882, + 0.012395162135362625, + 0.005384004209190607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07899297028779984 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007279845303855836 + ], + "opposite_logprob_tokens": [ + -7.250728130340576 + ], + "logprob_reward": -0.0007279845303855836, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007279845303855836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13810960948467255, + 0.11166204512119293, + 0.13125719130039215, + 0.09518583863973618, + 0.10618507117033005, + 0.07304619252681732, + 0.08991041779518127, + 0.09329409897327423, + 0.11187626421451569, + 0.049473248422145844 + ], + "credit_map": [ + 0.012702585197985172, + 0.016241680830717087, + 0.013878894969820976, + 0.017661405727267265, + 0.019107429310679436, + 0.014507665298879147, + 0.015850597992539406, + 0.02815934643149376, + 0.0220467709004879, + 0.01357240043580532, + 0.02413558028638363, + 0.007862593047320843, + 0.016049958765506744, + 0.009468551725149155, + 0.0075490642338991165, + 0.0109771229326725, + 0.01325707696378231, + 0.013246862217783928, + 0.014338440261781216, + 0.019929200410842896, + 0.016346922144293785, + 0.02365671843290329, + 0.01624676026403904, + 0.014235207810997963, + 0.013617639429867268, + 0.015078963711857796, + 0.01765456236898899, + 0.011031064204871655, + 0.00991992000490427, + 0.009882821701467037, + 0.009913705289363861, + 0.008087161928415298, + 0.01271736714988947, + 0.015186889097094536, + 0.011920654214918613, + 0.010595076717436314, + 0.013680326752364635, + 0.016157690435647964, + 0.012292916886508465, + 0.013634142465889454, + 0.006786084733903408, + 0.007903830148279667, + 0.007021992467343807, + 0.011684266850352287, + 0.010631956160068512, + 0.010662199929356575, + 0.010261179879307747, + 0.008094683289527893, + 0.008310050703585148, + 0.01362911518663168, + 0.00995927955955267, + 0.02388067916035652, + 0.008004647679626942, + 0.007004528306424618, + 0.012758448719978333, + 0.00636367266997695, + 0.011246085166931152, + 0.009274263866245747, + 0.008216015994548798, + 0.015729546546936035, + 0.013900761492550373, + 0.010431320406496525, + 0.014787094667553902, + 0.009709009900689125, + 0.010316187515854836, + 0.01368078961968422, + 0.011439898051321507, + 0.01960703544318676, + 0.014029868878424168, + 0.013498707674443722, + 0.020091582089662552, + 0.0092121921479702, + 0.004488498903810978, + 0.006235496141016483, + 0.005605948623269796, + 0.00864452961832285, + 0.005052895750850439, + 0.005716724321246147, + 0.009320481680333614, + 0.004408672451972961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0007279845303855836 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060391440987587 + ], + "opposite_logprob_tokens": [ + -4.1410603523254395 + ], + "logprob_reward": -0.016060391440987587, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060391440987587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11240115761756897, + 0.1045278012752533, + 0.09355007857084274, + 0.05404697731137276, + 0.11949606239795685, + 0.06870777904987335, + 0.11801058053970337, + 0.18593533337116241, + 0.09089429676532745, + 0.05242997407913208 + ], + "credit_map": [ + 0.010667253285646439, + 0.0137046929448843, + 0.011420812457799911, + 0.014982183463871479, + 0.011583298444747925, + 0.014016888104379177, + 0.017274336889386177, + 0.018751690164208412, + 0.023547936230897903, + 0.013854584656655788, + 0.014860428869724274, + 0.008608353324234486, + 0.009410091675817966, + 0.013008443638682365, + 0.006763280835002661, + 0.014474684372544289, + 0.008421642705798149, + 0.01211647130548954, + 0.011894461698830128, + 0.0153926070779562, + 0.012747582979500294, + 0.012243335135281086, + 0.011356458067893982, + 0.009377514012157917, + 0.00669654319062829, + 0.0056547862477600574, + 0.0070562660694122314, + 0.006392349023371935, + 0.005302627105265856, + 0.009296484291553497, + 0.0066909100860357285, + 0.0069570112973451614, + 0.01735525019466877, + 0.011231847107410431, + 0.01741628348827362, + 0.012610559351742268, + 0.008192990906536579, + 0.021239858120679855, + 0.013133476488292217, + 0.01831580512225628, + 0.006971742492169142, + 0.00749608501791954, + 0.006300697568804026, + 0.012615394778549671, + 0.01024964265525341, + 0.006039674859493971, + 0.013071638531982899, + 0.005962902680039406, + 0.013992920517921448, + 0.012997695244848728, + 0.011337070725858212, + 0.019880397245287895, + 0.01093993242830038, + 0.014630597084760666, + 0.023323820903897285, + 0.010908138938248158, + 0.01685291714966297, + 0.02944835089147091, + 0.01751699298620224, + 0.03835655748844147, + 0.022059274837374687, + 0.016008218750357628, + 0.03357313200831413, + 0.012119889259338379, + 0.006889636628329754, + 0.013029715046286583, + 0.010424143634736538, + 0.022291401401162148, + 0.010925850830972195, + 0.006704962346702814, + 0.015235762111842632, + 0.005392830818891525, + 0.004458693787455559, + 0.008031276054680347, + 0.0048098573461174965, + 0.00837553571909666, + 0.006097937934100628, + 0.005837226286530495, + 0.010033018887043, + 0.00478642713278532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.016060391440987587 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014182253740727901 + ], + "opposite_logprob_tokens": [ + -4.264182090759277 + ], + "logprob_reward": -0.014182253740727901, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014182253740727901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1159457415342331, + 0.05094342306256294, + 0.04959166795015335, + 0.045436009764671326, + 0.22437836229801178, + 0.08166520297527313, + 0.11877718567848206, + 0.10028506815433502, + 0.13646377623081207, + 0.07651352137327194 + ], + "credit_map": [ + 0.010653230361640453, + 0.012665635906159878, + 0.01596948318183422, + 0.015547734685242176, + 0.017036108300089836, + 0.0127408467233181, + 0.012084398418664932, + 0.019248303025960922, + 0.009756037965416908, + 0.00738740898668766, + 0.013895777054131031, + 0.0033203919883817434, + 0.00453559122979641, + 0.0038431400898844004, + 0.0029838429763913155, + 0.005221235100179911, + 0.004682807717472315, + 0.005408796481788158, + 0.005149831995368004, + 0.006692315451800823, + 0.006601484026759863, + 0.009637211449444294, + 0.0057389275170862675, + 0.005680294241756201, + 0.005447326693683863, + 0.006513912230730057, + 0.0054074618965387344, + 0.0069216652773320675, + 0.004957462660968304, + 0.0057001374661922455, + 0.005987643729895353, + 0.004500399809330702, + 0.03270503878593445, + 0.023182669654488564, + 0.029667826369404793, + 0.022212667390704155, + 0.018156765028834343, + 0.03990422189235687, + 0.023456236347556114, + 0.03509293496608734, + 0.008616937324404716, + 0.009481233544647694, + 0.008695066906511784, + 0.01335964072495699, + 0.00943137239664793, + 0.010366897098720074, + 0.010798277333378792, + 0.01091577485203743, + 0.012067868374288082, + 0.014002273790538311, + 0.010298915207386017, + 0.018558086827397346, + 0.010451637208461761, + 0.017388436943292618, + 0.023465732112526894, + 0.012544232420623302, + 0.013095344416797161, + 0.011831889860332012, + 0.00929278414696455, + 0.019830917939543724, + 0.01232664193958044, + 0.009474932216107845, + 0.014396493323147297, + 0.010036066174507141, + 0.012613029219210148, + 0.013823195360600948, + 0.01435698103159666, + 0.02404814399778843, + 0.01787496916949749, + 0.016277652233839035, + 0.026918819174170494, + 0.010550986044108868, + 0.0076298899948596954, + 0.011636652052402496, + 0.007551305461674929, + 0.008741218596696854, + 0.011576640419661999, + 0.008157978765666485, + 0.013359127566218376, + 0.007860704325139523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.014182253740727901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a bird's-eye view, a static extreme close-up captures a delicate, shimmering film of soapy water spreading across the smooth, dark surface of a polished granite countertop. Under soft, diffuse studio light, the film's surface is a captivating swirl of iridescent pink, cyan, and gold that slowly shifts and flows. The film thins noticeably, the colors swirling at an increasingly rapid pace until a breaking point appears near the center. Instantly, the fragile membrane fails, catastrophically collapsing inward from all sides as surface tension breaks down. The liquid coalesces, rushing swiftly and decisively toward the center in a chaotic surge that rapidly organizes and pulls the water into a nearly perfectly circular shape. The movement ceases almost as quickly as it began, leaving a clean, transparent wet ring on the countertop's smooth surface, slightly thicker at the edges. A few tiny, lingering bubbles shimmer within the clear liquid, creating a serene and mesmerizing vision of microscopic physics.", + "seed": 9955, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_2089_seed_9955_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": -0.03421087600290775, + "vlm_logprob": -0.03421087600290775, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03421087600290775, + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09662594646215439 + ], + "opposite_logprob_tokens": [ + -2.596626043319702 + ], + "logprob_reward": -0.09662594646215439, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09662594646215439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12937460839748383, + 0.10982772707939148, + 0.06889250874519348, + 0.11747954040765762, + 0.14963819086551666, + 0.07443755120038986, + 0.07442961633205414, + 0.09113000333309174, + 0.14800430834293365, + 0.03678583726286888 + ], + "credit_map": [ + 0.01500219851732254, + 0.014506933279335499, + 0.019796118140220642, + 0.017607318237423897, + 0.017695235088467598, + 0.015453645959496498, + 0.014995720237493515, + 0.014317444525659084, + 0.02518734522163868, + 0.009550055488944054, + 0.01952224411070347, + 0.005686141084879637, + 0.02077627368271351, + 0.010349738411605358, + 0.008725674822926521, + 0.010030270554125309, + 0.0063534569926559925, + 0.014348086901009083, + 0.006137478165328503, + 0.006857531610876322, + 0.0064891669899225235, + 0.010377512313425541, + 0.004946667235344648, + 0.01338261179625988, + 0.01277943141758442, + 0.018055101856589317, + 0.015838082879781723, + 0.01005938183516264, + 0.017284289002418518, + 0.017628170549869537, + 0.01570679247379303, + 0.010128304362297058, + 0.01532917469739914, + 0.017633624374866486, + 0.021825257688760757, + 0.016635216772556305, + 0.014576036483049393, + 0.015892047435045242, + 0.028400147333741188, + 0.019346706569194794, + 0.009354187175631523, + 0.005670921877026558, + 0.01112839300185442, + 0.006445165257900953, + 0.010303810238838196, + 0.01617611013352871, + 0.007098977454006672, + 0.00825999490916729, + 0.009288924746215343, + 0.00896652601659298, + 0.01142936758697033, + 0.0115541722625494, + 0.009576404467225075, + 0.008353044278919697, + 0.009458930231630802, + 0.005802251398563385, + 0.01745527796447277, + 0.006429191213101149, + 0.0176417026668787, + 0.006465922575443983, + 0.010156079195439816, + 0.012458170764148235, + 0.013094336725771427, + 0.007429329678416252, + 0.017396915704011917, + 0.016997231170535088, + 0.01417192816734314, + 0.018216870725154877, + 0.02826767787337303, + 0.02049274742603302, + 0.017573723569512367, + 0.014887227676808834, + 0.0040625520050525665, + 0.004154387395828962, + 0.006206061225384474, + 0.003622166346758604, + 0.004245974589139223, + 0.005870470777153969, + 0.0038616303354501724, + 0.00476260157302022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.09662594646215439 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496174469590187 + ], + "opposite_logprob_tokens": [ + -2.929961681365967 + ], + "logprob_reward": -0.05496174469590187, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496174469590187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07771196961402893, + 0.051472339779138565, + 0.029250921681523323, + 0.03929513320326805, + 0.3275454342365265, + 0.041526924818754196, + 0.04345712065696716, + 0.1480521410703659, + 0.19906039535999298, + 0.04262755438685417 + ], + "credit_map": [ + 0.00893922708928585, + 0.010115746408700943, + 0.013519917614758015, + 0.007458188105374575, + 0.008393912576138973, + 0.009454400278627872, + 0.011807535775005817, + 0.008023050613701344, + 0.010159267112612724, + 0.0033235030714422464, + 0.013504359871149063, + 0.002986752660945058, + 0.004651273600757122, + 0.007261027116328478, + 0.004772733431309462, + 0.00481342663988471, + 0.0031900822650641203, + 0.0036369245499372482, + 0.0036696589086204767, + 0.003577879862859845, + 0.002816330175846815, + 0.004422068130224943, + 0.0029110079631209373, + 0.005026971455663443, + 0.004620806314051151, + 0.00689521711319685, + 0.005224654451012611, + 0.003131403122097254, + 0.005173177924007177, + 0.004070267081260681, + 0.0073879859410226345, + 0.002791623817756772, + 0.03214838355779648, + 0.04070919752120972, + 0.049068596214056015, + 0.030598457902669907, + 0.03826681151986122, + 0.027519170194864273, + 0.05674545094370842, + 0.05248940736055374, + 0.005776610225439072, + 0.005312562454491854, + 0.006076985504478216, + 0.003146178787574172, + 0.006778866518288851, + 0.004546490963548422, + 0.005321659613400698, + 0.004567570984363556, + 0.004405935760587454, + 0.003680437570437789, + 0.007381185423582792, + 0.005718733184039593, + 0.005643186625093222, + 0.007756816688925028, + 0.0050461022183299065, + 0.0038247264456003904, + 0.017210904508829117, + 0.012609465047717094, + 0.026737211272120476, + 0.010386415757238865, + 0.0191508699208498, + 0.02808799035847187, + 0.018142415210604668, + 0.015726890414953232, + 0.023215817287564278, + 0.0198353361338377, + 0.021973630413413048, + 0.02021915279328823, + 0.03662493824958801, + 0.032885875552892685, + 0.022438399493694305, + 0.021867254748940468, + 0.005046309903264046, + 0.006416607648134232, + 0.003805654589086771, + 0.00620306096971035, + 0.0055144177749753, + 0.006059514824301004, + 0.006142690312117338, + 0.003439300460740924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.05496174469590187 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103217739611864 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.004103217739611864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103217739611864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20091275870800018, + 0.1295652538537979, + 0.06532786786556244, + 0.07378841936588287, + 0.18393287062644958, + 0.056783560663461685, + 0.05579991266131401, + 0.1577211171388626, + 0.04052399843931198, + 0.035644274204969406 + ], + "credit_map": [ + 0.03784192353487015, + 0.02469240501523018, + 0.031493812799453735, + 0.0169941708445549, + 0.018642812967300415, + 0.02009911648929119, + 0.031211627647280693, + 0.01993689313530922, + 0.041066743433475494, + 0.007534284610301256, + 0.01499178446829319, + 0.004970000125467777, + 0.009451650083065033, + 0.009525028057396412, + 0.00554382661357522, + 0.036481939256191254, + 0.0061753783375024796, + 0.00796138122677803, + 0.008568130433559418, + 0.009246466681361198, + 0.008136212825775146, + 0.009553777985274792, + 0.008905351161956787, + 0.006781165953725576, + 0.007480092812329531, + 0.009528507478535175, + 0.0179405827075243, + 0.005858393851667643, + 0.008651520125567913, + 0.006951618939638138, + 0.011764231137931347, + 0.0056134783662855625, + 0.018013279885053635, + 0.01977579854428768, + 0.03642430901527405, + 0.01589738205075264, + 0.021635286509990692, + 0.01733550801873207, + 0.02783043496310711, + 0.027020875364542007, + 0.007619173731654882, + 0.005338379181921482, + 0.007775670383125544, + 0.004858536645770073, + 0.009516500867903233, + 0.009012952446937561, + 0.006390816066414118, + 0.0062715294770896435, + 0.005804914981126785, + 0.005737569648772478, + 0.008082741871476173, + 0.008931634947657585, + 0.006675129756331444, + 0.008671578019857407, + 0.00703533086925745, + 0.00486100884154439, + 0.017637690529227257, + 0.015595161356031895, + 0.01779589243233204, + 0.008788283914327621, + 0.04062247276306152, + 0.01591092348098755, + 0.016357289627194405, + 0.025013402104377747, + 0.004671298898756504, + 0.004655216820538044, + 0.004320956766605377, + 0.005325559061020613, + 0.005598521325737238, + 0.006888127885758877, + 0.005196211859583855, + 0.0038681041914969683, + 0.0034431724343448877, + 0.004333171062171459, + 0.0036243293434381485, + 0.004180415999144316, + 0.004317541606724262, + 0.006489374674856663, + 0.005715455859899521, + 0.0035408150870352983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.004103217739611864 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13501368463039398, + 0.07899785786867142, + 0.05367623642086983, + 0.06250215321779251, + 0.15420332551002502, + 0.05573176592588425, + 0.08063948899507523, + 0.13473425805568695, + 0.18660999834537506, + 0.05789122357964516 + ], + "credit_map": [ + 0.0165624488145113, + 0.012764710932970047, + 0.012240506708621979, + 0.018267540261149406, + 0.015550974756479263, + 0.014233378693461418, + 0.03263905271887779, + 0.012755071744322777, + 0.02300078049302101, + 0.004711163695901632, + 0.006596282124519348, + 0.0037869925145059824, + 0.004364007618278265, + 0.004856104962527752, + 0.007446717005223036, + 0.024235809221863747, + 0.007384358439594507, + 0.00547225633636117, + 0.010071160271763802, + 0.006981394253671169, + 0.004956461954861879, + 0.0069852606393396854, + 0.005616344511508942, + 0.006208997685462236, + 0.006300094071775675, + 0.008246863260865211, + 0.012571512721478939, + 0.004427438136190176, + 0.009402877651154995, + 0.004724850878119469, + 0.011981590650975704, + 0.004846926778554916, + 0.021164216101169586, + 0.01236297283321619, + 0.03206227347254753, + 0.013103031553328037, + 0.014799244701862335, + 0.013994792476296425, + 0.024288451299071312, + 0.022428352385759354, + 0.006200067233294249, + 0.0068010310642421246, + 0.007221524603664875, + 0.008172466419637203, + 0.006169239524751902, + 0.00852083507925272, + 0.007135217078030109, + 0.005511389579623938, + 0.008756916038691998, + 0.00721993250772357, + 0.012395533733069897, + 0.008869826793670654, + 0.013544773682951927, + 0.01146012544631958, + 0.010568580590188503, + 0.00782379787415266, + 0.018688471987843513, + 0.014348375611007214, + 0.018930403515696526, + 0.009647444821894169, + 0.02851276658475399, + 0.012866601347923279, + 0.015168911777436733, + 0.016571270301938057, + 0.019881239160895348, + 0.02214963175356388, + 0.030436396598815918, + 0.0179036483168602, + 0.023894773796200752, + 0.03412267193198204, + 0.01996418833732605, + 0.01825745776295662, + 0.0057183848693966866, + 0.008161778561770916, + 0.00610237754881382, + 0.00708494707942009, + 0.009257460944354534, + 0.006172105669975281, + 0.009522504173219204, + 0.005871662404388189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.006734057795256376 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008629413321614265 + ], + "opposite_logprob_tokens": [ + -4.758629322052002 + ], + "logprob_reward": -0.008629413321614265, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008629413321614265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12150590866804123, + 0.11914330720901489, + 0.033081844449043274, + 0.044182728976011276, + 0.12047121673822403, + 0.06697451323270798, + 0.21035777032375336, + 0.07914505898952484, + 0.13631954789161682, + 0.06881812959909439 + ], + "credit_map": [ + 0.012230112217366695, + 0.011903495527803898, + 0.016998326405882835, + 0.015071393921971321, + 0.018099477514624596, + 0.011264043860137463, + 0.023591414093971252, + 0.012347673997282982, + 0.0394233763217926, + 0.006899416446685791, + 0.009958709590137005, + 0.004364497493952513, + 0.00737230246886611, + 0.007547127082943916, + 0.005723840557038784, + 0.03785404935479164, + 0.0031675926875323057, + 0.0037131644785404205, + 0.004878424108028412, + 0.002728820778429508, + 0.0027223373763263226, + 0.005578112322837114, + 0.003094441955909133, + 0.007198951207101345, + 0.00408860482275486, + 0.004752382170408964, + 0.00635077990591526, + 0.004296055994927883, + 0.007673921063542366, + 0.006440902128815651, + 0.006407653447240591, + 0.00417243642732501, + 0.010258322581648827, + 0.015216195024549961, + 0.017986668273806572, + 0.012004915624856949, + 0.015110168606042862, + 0.012288236059248447, + 0.02069736272096634, + 0.016909364610910416, + 0.009668117389082909, + 0.007089101243764162, + 0.011403614655137062, + 0.007185255642980337, + 0.009141849353909492, + 0.008103596977889538, + 0.007327747996896505, + 0.007055238354951143, + 0.018551161512732506, + 0.024859286844730377, + 0.02559441514313221, + 0.04107632488012314, + 0.02974221110343933, + 0.02379748225212097, + 0.029446467757225037, + 0.017290430143475533, + 0.011356142349541187, + 0.008084102533757687, + 0.010436910204589367, + 0.006347951479256153, + 0.01596670225262642, + 0.009418642148375511, + 0.009531457908451557, + 0.008003159426152706, + 0.015172751620411873, + 0.018045173957943916, + 0.009163036942481995, + 0.014977985993027687, + 0.03507803753018379, + 0.013731438666582108, + 0.018310025334358215, + 0.011841107159852982, + 0.007630160544067621, + 0.010288365185260773, + 0.006594122853130102, + 0.007515980862081051, + 0.012773633934557438, + 0.00783228874206543, + 0.010341164655983448, + 0.005842413753271103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.008629413321614265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a bird's-eye view, a static extreme close-up captures a delicate, shimmering film of soapy water spreading across the smooth, dark surface of a polished granite countertop. Under soft, diffuse studio light, the film's surface is a captivating swirl of iridescent pink, cyan, and gold that slowly shifts and flows. The film thins noticeably, the colors swirling at an increasingly rapid pace until a breaking point appears near the center. Instantly, the fragile membrane fails, catastrophically collapsing inward from all sides as surface tension breaks down. The liquid coalesces, rushing swiftly and decisively toward the center in a chaotic surge that rapidly organizes and pulls the water into a nearly perfectly circular shape. The movement ceases almost as quickly as it began, leaving a clean, transparent wet ring on the countertop's smooth surface, slightly thicker at the edges. A few tiny, lingering bubbles shimmer within the clear liquid, creating a serene and mesmerizing vision of microscopic physics.", + "seed": 9956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_12/index_2089_seed_9956_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": -0.030687269882764667, + "vlm_logprob": -0.030687269882764667, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.030687269882764667, + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037334591150283813 + ], + "opposite_logprob_tokens": [ + -3.537334680557251 + ], + "logprob_reward": -0.037334591150283813, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.037334591150283813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03907334432005882, + 0.030026836320757866, + 0.038220252841711044, + 0.7693868279457092, + 0.01686444878578186, + 0.02595115639269352, + 0.02945033833384514, + 0.025718923658132553, + 0.01498839259147644, + 0.010319430381059647 + ], + "credit_map": [ + 0.004213025793433189, + 0.004652475472539663, + 0.006496049929410219, + 0.004455319605767727, + 0.00339119671843946, + 0.004055712837725878, + 0.0059427847154438496, + 0.0058667827397584915, + 0.006808499339967966, + 0.003442126791924238, + 0.005457499995827675, + 0.002699981676414609, + 0.0028980234637856483, + 0.0022924961522221565, + 0.0025371606461703777, + 0.0038910498842597008, + 0.0026235724799335003, + 0.006971548777073622, + 0.0032505900599062443, + 0.0035261616576462984, + 0.002666337415575981, + 0.009000656194984913, + 0.0032224960159510374, + 0.006958894897252321, + 0.06870128959417343, + 0.2644001841545105, + 0.0735698714852333, + 0.024253949522972107, + 0.24712951481342316, + 0.029276825487613678, + 0.045683592557907104, + 0.016371658071875572, + 0.002390567446127534, + 0.0014708389062434435, + 0.0026249682996422052, + 0.0017581862630322576, + 0.0026578910183161497, + 0.0015944029437378049, + 0.002802440430969, + 0.0015651558060199022, + 0.00284001138061285, + 0.0029730896931141615, + 0.005341800395399332, + 0.0037565790116786957, + 0.0028725271113216877, + 0.002585014095529914, + 0.0033180387690663338, + 0.002264098497107625, + 0.003658764064311981, + 0.003041765885427594, + 0.0028145131655037403, + 0.003942770883440971, + 0.004500143695622683, + 0.005062707234174013, + 0.0035264440812170506, + 0.0029032311867922544, + 0.002408087719231844, + 0.003040290903300047, + 0.0017798158805817366, + 0.00236368621699512, + 0.00781425554305315, + 0.0018598862225189805, + 0.00399827491492033, + 0.0024546305648982525, + 0.001534924958832562, + 0.0016774405958130956, + 0.0014197282725945115, + 0.002238108078017831, + 0.002853491110727191, + 0.0011257623555138707, + 0.0022987998090684414, + 0.0018401375273242593, + 0.0008643323672004044, + 0.0010958821512758732, + 0.000887219503056258, + 0.0014238571748137474, + 0.002751304069533944, + 0.0008969403570517898, + 0.001407728879712522, + 0.0009921668097376823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.037334591150283813 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031463205814362 + ], + "opposite_logprob_tokens": [ + -2.3503146171569824 + ], + "logprob_reward": -0.10031463205814362, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031463205814362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05005381256341934, + 0.028343962505459785, + 0.012548972852528095, + 0.8160460591316223, + 0.01562674343585968, + 0.009214740246534348, + 0.031479958444833755, + 0.013508918695151806, + 0.011528180912137032, + 0.011648609302937984 + ], + "credit_map": [ + 0.005772892385721207, + 0.006668934132903814, + 0.006655966397374868, + 0.00734398839995265, + 0.005344879347831011, + 0.007703698240220547, + 0.005565892439335585, + 0.004997566808015108, + 0.004920424893498421, + 0.0034984464291483164, + 0.0057008047588169575, + 0.003026157384738326, + 0.0038686024490743876, + 0.002083131344988942, + 0.0022505398374050856, + 0.002995859133079648, + 0.001007087412290275, + 0.0015727877616882324, + 0.001386728836223483, + 0.0015639212215319276, + 0.0011458545923233032, + 0.002306262729689479, + 0.00135329132899642, + 0.002213041065260768, + 0.07589085400104523, + 0.28573232889175415, + 0.07875818014144897, + 0.027515674009919167, + 0.24794207513332367, + 0.03202304616570473, + 0.049035556614398956, + 0.019148409366607666, + 0.001521838130429387, + 0.0014452326577156782, + 0.0023945095017552376, + 0.001587767736054957, + 0.0021336697973310947, + 0.001703309128060937, + 0.003332653781399131, + 0.0015077651478350163, + 0.0012564599746838212, + 0.0012380186235532165, + 0.0016247035237029195, + 0.0013513326412066817, + 0.001007167506031692, + 0.000759438204113394, + 0.0011649797670543194, + 0.0008126407628878951, + 0.003220183774828911, + 0.0037416163831949234, + 0.0026170052587985992, + 0.0033527263440191746, + 0.007004431448876858, + 0.004738041665405035, + 0.0036939505953341722, + 0.0031120050698518753, + 0.0015031058574095368, + 0.0014589151833206415, + 0.001280047232285142, + 0.0012372261844575405, + 0.0036489253398030996, + 0.0012285111006349325, + 0.0017936326330527663, + 0.0013585565611720085, + 0.0012286327546462417, + 0.0015219347551465034, + 0.0012340537505224347, + 0.0012954488629475236, + 0.002299478743225336, + 0.0014316145097836852, + 0.001133250305429101, + 0.0013837687438353896, + 0.0011305940570309758, + 0.0017337629105895758, + 0.0014427413698285818, + 0.001412874204106629, + 0.001900523086078465, + 0.0014000412775203586, + 0.001324298675172031, + 0.001303775585256517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.10031463205814362 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004444326623342931 + ], + "opposite_logprob_tokens": [ + -7.750444412231445 + ], + "logprob_reward": -0.0004444326623342931, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004444326623342931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18387523293495178, + 0.041904762387275696, + 0.09039930999279022, + 0.40123963356018066, + 0.04081955924630165, + 0.06765913218259811, + 0.04770561307668686, + 0.03550620377063751, + 0.05518088862299919, + 0.0357096865773201 + ], + "credit_map": [ + 0.020165415480732918, + 0.024270279332995415, + 0.029643123969435692, + 0.019260691478848457, + 0.02118220552802086, + 0.024632541462779045, + 0.025350511074066162, + 0.019370462745428085, + 0.004963121842592955, + 0.004717693664133549, + 0.008587400428950787, + 0.004146168008446693, + 0.006482807453721762, + 0.004506073426455259, + 0.004020270891487598, + 0.004481229931116104, + 0.009176520630717278, + 0.013827696442604065, + 0.010213383473455906, + 0.009442353621125221, + 0.0071791112422943115, + 0.012089273892343044, + 0.008710836991667747, + 0.0197601281106472, + 0.035627324134111404, + 0.13603731989860535, + 0.03852378949522972, + 0.012569664046168327, + 0.12960462272167206, + 0.015803975984454155, + 0.023942040279507637, + 0.00913088209927082, + 0.004925262648612261, + 0.004582786466926336, + 0.006411134265363216, + 0.0038856633473187685, + 0.005284755490720272, + 0.0035741638857871294, + 0.007744231261312962, + 0.004411563742905855, + 0.006800755858421326, + 0.00823158584535122, + 0.008659820072352886, + 0.010019073262810707, + 0.008175949566066265, + 0.006923631764948368, + 0.010701645165681839, + 0.008146669715642929, + 0.006369752809405327, + 0.004338249564170837, + 0.006075382698327303, + 0.005636400077491999, + 0.004838141147047281, + 0.00869324617087841, + 0.006582699716091156, + 0.005171740427613258, + 0.0037660642992705107, + 0.0036645769141614437, + 0.003155810059979558, + 0.00434158043935895, + 0.00792714674025774, + 0.003778897225856781, + 0.004445585887879133, + 0.0044265431351959705, + 0.005016791168600321, + 0.007623495068401098, + 0.004671459086239338, + 0.007527326233685017, + 0.012844415381550789, + 0.005187289789319038, + 0.006682934705168009, + 0.005627174861729145, + 0.0032119438983500004, + 0.004526951815932989, + 0.005117122549563646, + 0.004207179881632328, + 0.007420397829264402, + 0.0035613931249827147, + 0.003874292131513357, + 0.003790403949096799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0004444326623342931 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028166405390948057 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.0028166405390948057, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028166405390948057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10101641714572906, + 0.07823769003152847, + 0.10803773999214172, + 0.4468199908733368, + 0.04915686696767807, + 0.05020180717110634, + 0.052387114614248276, + 0.03698926791548729, + 0.05671834945678711, + 0.020434843376278877 + ], + "credit_map": [ + 0.009664795361459255, + 0.01335544791072607, + 0.013857533223927021, + 0.009135029278695583, + 0.014091749675571918, + 0.017679408192634583, + 0.013011276721954346, + 0.010221170261502266, + 0.00857175886631012, + 0.008440078236162663, + 0.01211534533649683, + 0.011233698576688766, + 0.015146438963711262, + 0.007224107161164284, + 0.008120466023683548, + 0.00738579872995615, + 0.011380566284060478, + 0.014774738810956478, + 0.011255635879933834, + 0.011563294567167759, + 0.008811513893306255, + 0.012247969396412373, + 0.011535351164638996, + 0.026468662545084953, + 0.04261454939842224, + 0.15522147715091705, + 0.04361111670732498, + 0.019725611433386803, + 0.11999492347240448, + 0.021991480141878128, + 0.027277421206235886, + 0.016383429989218712, + 0.00669678533449769, + 0.004918771330267191, + 0.006718148943036795, + 0.007072214037179947, + 0.006838577333837748, + 0.004645742010325193, + 0.008192826993763447, + 0.00407380098477006, + 0.0040283408015966415, + 0.0068148355931043625, + 0.007174999453127384, + 0.007014736998826265, + 0.007159681990742683, + 0.005413916427642107, + 0.008276316337287426, + 0.004318980034440756, + 0.006082866340875626, + 0.005643150769174099, + 0.0051770745776593685, + 0.006296460051089525, + 0.0087931789457798, + 0.009351097978651524, + 0.005737750791013241, + 0.005305534228682518, + 0.004353058058768511, + 0.004015322308987379, + 0.003696318017318845, + 0.0040192934684455395, + 0.008757567033171654, + 0.004268627148121595, + 0.004110473208129406, + 0.0037686112336814404, + 0.004877310711890459, + 0.005804079584777355, + 0.00356249394826591, + 0.00945402029901743, + 0.011616520583629608, + 0.005769714713096619, + 0.008056600578129292, + 0.00757761113345623, + 0.0028471960686147213, + 0.003074999200180173, + 0.002848018193617463, + 0.002163239289075136, + 0.003553587244823575, + 0.0015230011194944382, + 0.0024828105233609676, + 0.0019419927848502994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0028166405390948057 + }, + { + "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 bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene an extreme close-up of a soapy water film on a polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the film display iridescent pink, cyan, and gold colors that swirl and flow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the film catastrophically collapse inward, causing the liquid to rush toward the center and form a nearly perfectly circular shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final state show a clean, transparent wet ring with tiny lingering bubbles on the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012526053003966808 + ], + "opposite_logprob_tokens": [ + -4.387526035308838 + ], + "logprob_reward": -0.012526053003966808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012526053003966808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051333483308553696, + 0.034717388451099396, + 0.040441956371068954, + 0.6168830990791321, + 0.052536748349666595, + 0.04156867414712906, + 0.05116082355380058, + 0.0352453850209713, + 0.04102011024951935, + 0.03509235754609108 + ], + "credit_map": [ + 0.006621218286454678, + 0.006543265655636787, + 0.00824505090713501, + 0.006690313573926687, + 0.005663461051881313, + 0.004477063659578562, + 0.00785654503852129, + 0.005236563738435507, + 0.004324209410697222, + 0.004112434573471546, + 0.006256558001041412, + 0.002843169029802084, + 0.0038429908454418182, + 0.004133082926273346, + 0.0028045561630278826, + 0.006400387268513441, + 0.004251479636877775, + 0.008710363879799843, + 0.005308634135872126, + 0.005002730526030064, + 0.004054384306073189, + 0.003775407327339053, + 0.0025981932412832975, + 0.00674076471477747, + 0.05832185223698616, + 0.21704798936843872, + 0.060047347098588943, + 0.021466270089149475, + 0.1825025975704193, + 0.02445308119058609, + 0.03778767213225365, + 0.015256267972290516, + 0.005211770534515381, + 0.004194409586489201, + 0.005476747173815966, + 0.0059285955503582954, + 0.00781190674751997, + 0.005348845385015011, + 0.014154285192489624, + 0.004410186782479286, + 0.004512020852416754, + 0.004459408111870289, + 0.005195363890379667, + 0.006719671189785004, + 0.006144573446363211, + 0.004518556408584118, + 0.005829241126775742, + 0.004189836792647839, + 0.005393884610384703, + 0.0055940099991858006, + 0.006617341656237841, + 0.00552342040464282, + 0.008774320594966412, + 0.007142916321754456, + 0.007513201795518398, + 0.004601729568094015, + 0.0035600438714027405, + 0.0047765388153493404, + 0.002783760894089937, + 0.0029150855261832476, + 0.00848117284476757, + 0.005280744284391403, + 0.003724576672539115, + 0.0037234623450785875, + 0.00486381258815527, + 0.004869057331234217, + 0.00449441047385335, + 0.004025678150355816, + 0.007924167439341545, + 0.00332720298320055, + 0.005812481045722961, + 0.00570330023765564, + 0.0030592752154916525, + 0.00413997657597065, + 0.004718077834695578, + 0.006082022096961737, + 0.005126483738422394, + 0.0029259922448545694, + 0.0045278072357177734, + 0.004512721206992865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.012526053003966808 + } + ], + "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/0000017/group_13/index_tensor([173, 173])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_tensor([173, 173])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb8c3039ea0e9fdd1acccc508e783542ae327ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_tensor([173, 173])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, rendered in a gritty, unadorned photorealistic style, captures the raw power of a massive Hummer H1 Alpha as the camera begins a slow, steady pullback. Painted a matte olive drab green and almost entirely caked in thick, wet mud, the vehicle's iconic boxy shape laboriously climbs a steep, treacherous slope. Its massive, deeply treaded tires spin with immense torque, relentlessly digging into the saturated terrain and violently propelling arcs of dark brown soil and small rocks backward. The entire chassis exerts itself, the suspension articulating the noticeably uneven ground. The scene is bathed in the cool, unfocused light of an overcast sky, highlighting the glistening dampness of the tumbling mud and the splattered dirt clinging to the vehicle's chassis. A dense wall of dark green forest foliage, slightly out of focus in the background, combines with the dominant palette of dark browns and greens to create an atmosphere of intense, primal struggle between a powerful machine and the rugged landscape. The camera's slow, backward movement continuously emphasizes the vehicle's immense scale and the immense difficulty of the ascent.", + "seed": 692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_692_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": -0.11346076666377485, + "vlm_logprob": -0.11346076666377485, + "vlm_yes_no_margin": 3.525000011920929, + "vlm_token_credit": -0.11346076666377485, + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1119246557354927 + ], + "opposite_logprob_tokens": [ + -2.486924648284912 + ], + "logprob_reward": -0.1119246557354927, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1119246557354927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23011662065982819, + 0.04822177439928055, + 0.03934856131672859, + 0.06184537336230278, + 0.05804004520177841, + 0.03734743222594261, + 0.25694191455841064, + 0.0377269871532917, + 0.0802464410662651, + 0.15016482770442963 + ], + "credit_map": [ + 0.031614601612091064, + 0.01884632371366024, + 0.0455559641122818, + 0.02408834919333458, + 0.01835816726088524, + 0.030314253643155098, + 0.023519093170762062, + 0.03781987354159355, + 0.003968722186982632, + 0.005118913482874632, + 0.006033687852323055, + 0.006192736327648163, + 0.007176450919359922, + 0.005379796028137207, + 0.009376635774970055, + 0.004974829498678446, + 0.0048939320258796215, + 0.0048363483510911465, + 0.005200006067752838, + 0.00639113737270236, + 0.004598923027515411, + 0.003787020454183221, + 0.005959582515060902, + 0.003681613365188241, + 0.009968404658138752, + 0.009719019755721092, + 0.007573865819722414, + 0.008069675415754318, + 0.006526560056954622, + 0.006554039660841227, + 0.006362684536725283, + 0.007071124389767647, + 0.016090963035821915, + 0.003548596752807498, + 0.01391774695366621, + 0.0068523213267326355, + 0.003939800430089235, + 0.00415082136169076, + 0.004902614746242762, + 0.004637176636606455, + 0.005524924024939537, + 0.004280252382159233, + 0.005935506895184517, + 0.0037659225054085255, + 0.004803229123353958, + 0.005729088094085455, + 0.0032488617580384016, + 0.004059647209942341, + 0.07912174612283707, + 0.019139472395181656, + 0.08505803346633911, + 0.014034378342330456, + 0.02460125833749771, + 0.013218509964644909, + 0.00954873114824295, + 0.012219774536788464, + 0.006508020684123039, + 0.004067745991051197, + 0.005331849679350853, + 0.004518052563071251, + 0.003620134200900793, + 0.003527643159031868, + 0.004366349894553423, + 0.005787189584225416, + 0.008467916399240494, + 0.02428927831351757, + 0.005661224480718374, + 0.004988234490156174, + 0.006863696034997702, + 0.004724704660475254, + 0.004322103690356016, + 0.020929288119077682, + 0.010818450711667538, + 0.00953271146863699, + 0.03084227815270424, + 0.015023892745375633, + 0.015546365641057491, + 0.01477344986051321, + 0.03446032479405403, + 0.019167345017194748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.1119246557354927 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00367668061517179 + ], + "opposite_logprob_tokens": [ + -5.628676891326904 + ], + "logprob_reward": -0.00367668061517179, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00367668061517179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11395104229450226, + 0.14635345339775085, + 0.05453365296125412, + 0.16696326434612274, + 0.05344107747077942, + 0.026154544204473495, + 0.13197089731693268, + 0.04276106879115105, + 0.10413122177124023, + 0.15973980724811554 + ], + "credit_map": [ + 0.010578677989542484, + 0.014172190800309181, + 0.01686083897948265, + 0.012006876058876514, + 0.009807994589209557, + 0.014842513017356396, + 0.023655293509364128, + 0.012026645243167877, + 0.012645707465708256, + 0.015619467943906784, + 0.011544561944901943, + 0.02136589027941227, + 0.024404842406511307, + 0.016820721328258514, + 0.026603229343891144, + 0.01734902895987034, + 0.007802019361406565, + 0.007854391820728779, + 0.005051470827311277, + 0.009304898791015148, + 0.00680667394772172, + 0.005683611147105694, + 0.007784660905599594, + 0.004245922435075045, + 0.02330830693244934, + 0.025279400870203972, + 0.02159344218671322, + 0.01954931579530239, + 0.015990832820534706, + 0.02058662846684456, + 0.02254762500524521, + 0.018107715994119644, + 0.01869089901447296, + 0.002409091917797923, + 0.018295854330062866, + 0.0031147953122854233, + 0.002910376526415348, + 0.003908159676939249, + 0.001916480134241283, + 0.002195421839132905, + 0.004919333849102259, + 0.002621864899992943, + 0.004869156517088413, + 0.0029560779221355915, + 0.002177760936319828, + 0.003473725402727723, + 0.0025600020308047533, + 0.0025766228791326284, + 0.04567134007811546, + 0.009980946779251099, + 0.05208094045519829, + 0.005822575651109219, + 0.0053087761625647545, + 0.0035955533385276794, + 0.0031652129255235195, + 0.0063455612398684025, + 0.008450026623904705, + 0.004978750366717577, + 0.006657930091023445, + 0.005581901874393225, + 0.003727290080860257, + 0.003997043706476688, + 0.0034629537258297205, + 0.005905171390622854, + 0.01197428535670042, + 0.02096516266465187, + 0.009941695258021355, + 0.007640022784471512, + 0.00898802001029253, + 0.01012346614152193, + 0.00979292206466198, + 0.024705640971660614, + 0.010462547652423382, + 0.005402586422860622, + 0.05501753091812134, + 0.009884064085781574, + 0.007966294884681702, + 0.007038648705929518, + 0.054183632135391235, + 0.009784513153135777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.00367668061517179 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214339666068554 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214339666068554, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214339666068554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20137859880924225, + 0.03639166057109833, + 0.05072099715471268, + 0.028189266100525856, + 0.1236150935292244, + 0.021321866661310196, + 0.3867797255516052, + 0.019463978707790375, + 0.024347539991140366, + 0.10779125243425369 + ], + "credit_map": [ + 0.02364320307970047, + 0.023086780682206154, + 0.031134657561779022, + 0.015566430054605007, + 0.022254416719079018, + 0.033563025295734406, + 0.02121674455702305, + 0.030913354828953743, + 0.0034269869793206453, + 0.0037345848977565765, + 0.004045406822115183, + 0.004403702914714813, + 0.00406425166875124, + 0.004091174341738224, + 0.005102994851768017, + 0.0075225625187158585, + 0.006651707924902439, + 0.00561955850571394, + 0.008362110704183578, + 0.005299957934767008, + 0.006660170387476683, + 0.006500215269625187, + 0.0071437605656683445, + 0.0044835228472948074, + 0.004252931568771601, + 0.003353139851242304, + 0.005046561360359192, + 0.0032468847930431366, + 0.0030530847143381834, + 0.0036655536387115717, + 0.003029878716915846, + 0.002541232155635953, + 0.04441423341631889, + 0.006035913247615099, + 0.04347620904445648, + 0.005724259652197361, + 0.005996783263981342, + 0.007068105041980743, + 0.005079575348645449, + 0.005820021498948336, + 0.003485203953459859, + 0.0016085620736703277, + 0.004199488554149866, + 0.0027638606261461973, + 0.0018419417319819331, + 0.003594109322875738, + 0.0017965944716706872, + 0.0020321074407547712, + 0.13244760036468506, + 0.030901707708835602, + 0.1525270789861679, + 0.017882799729704857, + 0.011932437308132648, + 0.010987693443894386, + 0.010549585334956646, + 0.019550854340195656, + 0.005372477229684591, + 0.00259946146979928, + 0.0027641698252409697, + 0.0020754060242325068, + 0.0014541682321578264, + 0.0016582051757723093, + 0.0014249574160203338, + 0.002115133684128523, + 0.0033419032115489244, + 0.004763931501656771, + 0.001934126135893166, + 0.002000294392928481, + 0.0027588196098804474, + 0.004170876462012529, + 0.0027739594224840403, + 0.00260363332927227, + 0.007317241746932268, + 0.0032427601981908083, + 0.03715725615620613, + 0.005033985711634159, + 0.0049765706062316895, + 0.0048142713494598866, + 0.03738173469901085, + 0.007867440581321716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.014214339666068554 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637804538011551 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637804538011551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637804538011551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038414597511291504, + 0.027212807908654213, + 0.02952704206109047, + 0.03465771675109863, + 0.046560801565647125, + 0.049958985298871994, + 0.5225611329078674, + 0.016836008056998253, + 0.05487445369362831, + 0.17939640581607819 + ], + "credit_map": [ + 0.004145361483097076, + 0.005793070420622826, + 0.006064142566174269, + 0.0036384863778948784, + 0.0031209238804876804, + 0.005463704466819763, + 0.006095427554100752, + 0.004093481227755547, + 0.0026831477880477905, + 0.0026509345043450594, + 0.004463043063879013, + 0.003475327044725418, + 0.0034496465232223272, + 0.003057822585105896, + 0.0033174478448927402, + 0.004115438088774681, + 0.0041069104336202145, + 0.002481191884726286, + 0.005042899399995804, + 0.0032738931477069855, + 0.002853741403669119, + 0.003867323277518153, + 0.004884457215666771, + 0.0030166234355419874, + 0.00548347644507885, + 0.0036466424353420734, + 0.00670145358890295, + 0.0031824782490730286, + 0.0030185021460056305, + 0.00460670655593276, + 0.0037257990334182978, + 0.004292660392820835, + 0.013336616568267345, + 0.0026033332105726004, + 0.012153715826570988, + 0.002972930669784546, + 0.002659834222868085, + 0.008193041197955608, + 0.0019961949437856674, + 0.002645134227350354, + 0.00935483630746603, + 0.005126631818711758, + 0.0117582306265831, + 0.007410976104438305, + 0.0034397973213344812, + 0.004432670306414366, + 0.004126857500523329, + 0.004308987408876419, + 0.1801537275314331, + 0.04130033403635025, + 0.20724302530288696, + 0.02328915148973465, + 0.015894711017608643, + 0.015219572931528091, + 0.013562769629061222, + 0.025897817686200142, + 0.004194484557956457, + 0.0017456518253311515, + 0.0031877809669822454, + 0.0017653618706390262, + 0.0012171382550150156, + 0.0015660148346796632, + 0.0015558269806206226, + 0.0016037487657740712, + 0.006297678221017122, + 0.015024234540760517, + 0.004909338429570198, + 0.004573548212647438, + 0.004434994887560606, + 0.004604663699865341, + 0.005331073421984911, + 0.009698922745883465, + 0.012307927943766117, + 0.005962963681668043, + 0.06337380409240723, + 0.007764675188809633, + 0.0076553407125175, + 0.007359602954238653, + 0.06233518198132515, + 0.01263690460473299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.008637804538011551 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42885035276412964 + ], + "opposite_logprob_tokens": [ + -1.0538504123687744 + ], + "logprob_reward": -0.42885035276412964, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42885035276412964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07040305435657501, + 0.024654170498251915, + 0.020596003159880638, + 0.01516647543758154, + 0.018496083095669746, + 0.05804206803441048, + 0.6905634999275208, + 0.02005784586071968, + 0.036442287266254425, + 0.04557842016220093 + ], + "credit_map": [ + 0.008361393585801125, + 0.005653392057865858, + 0.010790987871587276, + 0.006722001824527979, + 0.005923843011260033, + 0.01132708229124546, + 0.013372653163969517, + 0.008251703344285488, + 0.002369851339608431, + 0.00346006010659039, + 0.003084048628807068, + 0.003418539883568883, + 0.0031343591399490833, + 0.0027677929028868675, + 0.004417038522660732, + 0.002002483233809471, + 0.0035225071478635073, + 0.002036717254668474, + 0.002445878926664591, + 0.0031713717617094517, + 0.002249316545203328, + 0.002815317129716277, + 0.0026657776907086372, + 0.0016891187988221645, + 0.0022097378969192505, + 0.002458875300362706, + 0.001910615828819573, + 0.0016518707852810621, + 0.0014579427661374211, + 0.0013926144456490874, + 0.0022627480793744326, + 0.0018220721976831555, + 0.003991970792412758, + 0.001351747428998351, + 0.0029795500449836254, + 0.0021724519319832325, + 0.0015561063773930073, + 0.0020022850949317217, + 0.0021383697167038918, + 0.0023036031052470207, + 0.009965538047254086, + 0.0069540501572191715, + 0.012668217532336712, + 0.008245396427810192, + 0.004073342774063349, + 0.004582760389894247, + 0.005638347938656807, + 0.005914417095482349, + 0.23790720105171204, + 0.05470835044980049, + 0.27342960238456726, + 0.031103068962693214, + 0.021294480189681053, + 0.020128333941102028, + 0.018071765080094337, + 0.03392075002193451, + 0.003983055241405964, + 0.002429636660963297, + 0.0030009043402969837, + 0.002024469431489706, + 0.0016761255683377385, + 0.0024687510449439287, + 0.0019690385088324547, + 0.0025058689061552286, + 0.004380595870316029, + 0.010617238469421864, + 0.003195203375071287, + 0.0020875430200248957, + 0.0025533651933073997, + 0.0034466783981770277, + 0.00284445658326149, + 0.007317212410271168, + 0.004161690827459097, + 0.004339220467954874, + 0.011651560664176941, + 0.0032335487194359303, + 0.002735380781814456, + 0.0031025605276226997, + 0.011373843066394329, + 0.0049806153401732445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.42885035276412964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, rendered in a gritty, unadorned photorealistic style, captures the raw power of a massive Hummer H1 Alpha as the camera begins a slow, steady pullback. Painted a matte olive drab green and almost entirely caked in thick, wet mud, the vehicle's iconic boxy shape laboriously climbs a steep, treacherous slope. Its massive, deeply treaded tires spin with immense torque, relentlessly digging into the saturated terrain and violently propelling arcs of dark brown soil and small rocks backward. The entire chassis exerts itself, the suspension articulating the noticeably uneven ground. The scene is bathed in the cool, unfocused light of an overcast sky, highlighting the glistening dampness of the tumbling mud and the splattered dirt clinging to the vehicle's chassis. A dense wall of dark green forest foliage, slightly out of focus in the background, combines with the dominant palette of dark browns and greens to create an atmosphere of intense, primal struggle between a powerful machine and the rugged landscape. The camera's slow, backward movement continuously emphasizes the vehicle's immense scale and the immense difficulty of the ascent.", + "seed": 693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_693_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": -0.08486855356022716, + "vlm_logprob": -0.08486855356022716, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08486855356022716, + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2207242101430893 + ], + "opposite_logprob_tokens": [ + -1.7207242250442505 + ], + "logprob_reward": -0.2207242101430893, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2207242101430893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11307170987129211, + 0.032169368118047714, + 0.05494486168026924, + 0.43484827876091003, + 0.045709654688835144, + 0.03899112716317177, + 0.03891848400235176, + 0.16228275001049042, + 0.03631770238280296, + 0.04274604097008705 + ], + "credit_map": [ + 0.013416987843811512, + 0.016224712133407593, + 0.013824868947267532, + 0.015945596620440483, + 0.012912617065012455, + 0.016359589993953705, + 0.012862036004662514, + 0.01152530312538147, + 0.0036932046059519053, + 0.003184674773365259, + 0.002521494636312127, + 0.0025488256942480803, + 0.005719475913792849, + 0.00399670097976923, + 0.0046737417578697205, + 0.005831251386553049, + 0.008096718229353428, + 0.007649367209523916, + 0.004458262585103512, + 0.0044610160402953625, + 0.009210601449012756, + 0.006562155205756426, + 0.007962212897837162, + 0.006544529926031828, + 0.019169120118021965, + 0.010999745689332485, + 0.012901011854410172, + 0.006300742272287607, + 0.1695355772972107, + 0.09604228287935257, + 0.023101281374692917, + 0.09679852426052094, + 0.005426226183772087, + 0.0037320214323699474, + 0.00415327912196517, + 0.0051867044530808926, + 0.006669777445495129, + 0.004550254438072443, + 0.005158726125955582, + 0.01083266269415617, + 0.00467277504503727, + 0.0034357367549091578, + 0.0029063925612717867, + 0.003986583091318607, + 0.004693405237048864, + 0.007868706248700619, + 0.00449413014575839, + 0.006933400873094797, + 0.004306599497795105, + 0.004517620895057917, + 0.005189241375774145, + 0.0065938192419707775, + 0.004056411795318127, + 0.003962858580052853, + 0.004915805999189615, + 0.005376126617193222, + 0.02156945876777172, + 0.011075341142714024, + 0.019621100276708603, + 0.0244076419621706, + 0.01679699681699276, + 0.019118735566735268, + 0.03492092341184616, + 0.014772549271583557, + 0.005741295870393515, + 0.003124938579276204, + 0.0036795660853385925, + 0.006863793358206749, + 0.004465402103960514, + 0.0033942079171538353, + 0.005228950642049313, + 0.003819551784545183, + 0.006725459359586239, + 0.0027883967850357294, + 0.003989269956946373, + 0.005016372073441744, + 0.007310399319976568, + 0.004500063136219978, + 0.00831631850451231, + 0.0040997606702148914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.2207242101430893 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008708239533007145 + ], + "opposite_logprob_tokens": [ + -4.758708477020264 + ], + "logprob_reward": -0.008708239533007145, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008708239533007145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13685309886932373, + 0.14627213776111603, + 0.21705952286720276, + 0.08101971447467804, + 0.041712384670972824, + 0.04622691124677658, + 0.04612426832318306, + 0.07506559044122696, + 0.1296512931585312, + 0.08001510053873062 + ], + "credit_map": [ + 0.014378917403519154, + 0.01901957206428051, + 0.014058388769626617, + 0.01629551686346531, + 0.01793603040277958, + 0.021380867809057236, + 0.01777762547135353, + 0.01600615121424198, + 0.014178170822560787, + 0.016760222613811493, + 0.010780008509755135, + 0.01668670028448105, + 0.029423126950860023, + 0.0197671540081501, + 0.0175341609865427, + 0.021142570301890373, + 0.023855486884713173, + 0.017192795872688293, + 0.0120093347504735, + 0.019588176161050797, + 0.01977785862982273, + 0.03150147199630737, + 0.05055832490324974, + 0.04257604479789734, + 0.007021090481430292, + 0.006056778132915497, + 0.005592600908130407, + 0.0051919580437242985, + 0.01824689283967018, + 0.022606484591960907, + 0.008005905896425247, + 0.008297989144921303, + 0.003599767107516527, + 0.002224684925749898, + 0.003180096857249737, + 0.004335378296673298, + 0.006853069178760052, + 0.006448474712669849, + 0.004937523044645786, + 0.010133386589586735, + 0.004187142476439476, + 0.006084953434765339, + 0.004408621229231358, + 0.0056900219060480595, + 0.0072189811617136, + 0.005198899190872908, + 0.007701932918280363, + 0.005736352410167456, + 0.0061153327114880085, + 0.0037557664327323437, + 0.007388195488601923, + 0.006032124161720276, + 0.0032037529163062572, + 0.005306086968630552, + 0.007733159698545933, + 0.0065898471511900425, + 0.011260985396802425, + 0.008734888397157192, + 0.01146676391363144, + 0.009162571281194687, + 0.006783203687518835, + 0.010074149817228317, + 0.008526084013283253, + 0.0090569369494915, + 0.022070733830332756, + 0.012347426265478134, + 0.0092921769246459, + 0.0319395512342453, + 0.014661040157079697, + 0.01030073780566454, + 0.018620947375893593, + 0.010418664664030075, + 0.013202176429331303, + 0.004944975022226572, + 0.00913166906684637, + 0.012905946932733059, + 0.009295227937400341, + 0.00647411122918129, + 0.015742624178528786, + 0.008318359032273293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.008708239533007145 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806079789996147 + ], + "opposite_logprob_tokens": [ + -4.634806156158447 + ], + "logprob_reward": -0.009806079789996147, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806079789996147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13495688140392303, + 0.1736096739768982, + 0.07421430945396423, + 0.3290795385837555, + 0.04882800579071045, + 0.05727324262261391, + 0.07883070409297943, + 0.04915948212146759, + 0.028308497741818428, + 0.025739647448062897 + ], + "credit_map": [ + 0.017366701737046242, + 0.015234114602208138, + 0.019214067608118057, + 0.012558752670884132, + 0.01557490136474371, + 0.020066697150468826, + 0.012936015613377094, + 0.022005638107657433, + 0.013120411895215511, + 0.019684849306941032, + 0.011407939717173576, + 0.021771688014268875, + 0.02785937301814556, + 0.024662533774971962, + 0.023077812045812607, + 0.0320250540971756, + 0.009913323447108269, + 0.006994463037699461, + 0.004470966290682554, + 0.006033074110746384, + 0.01257671695202589, + 0.009204141795635223, + 0.014401611872017384, + 0.010620014742016792, + 0.017482157796621323, + 0.009510780684649944, + 0.014629852026700974, + 0.00802447646856308, + 0.12115791440010071, + 0.056424908339977264, + 0.017822187393903732, + 0.0840272530913353, + 0.004180990159511566, + 0.003439669730141759, + 0.005558650474995375, + 0.006559943314641714, + 0.007174775004386902, + 0.005321627017110586, + 0.0060050212778151035, + 0.010587328113615513, + 0.005749847274273634, + 0.005269988439977169, + 0.0048963762819767, + 0.005935773719102144, + 0.005214916076511145, + 0.008531861938536167, + 0.007228718139231205, + 0.014445764012634754, + 0.009541596285998821, + 0.0073038083501160145, + 0.014382460154592991, + 0.010295027866959572, + 0.007450190372765064, + 0.008328188210725784, + 0.012294428423047066, + 0.009234997443854809, + 0.006335737183690071, + 0.005859320517629385, + 0.006962829269468784, + 0.005409444682300091, + 0.0064818342216312885, + 0.005420457571744919, + 0.008642241358757019, + 0.0040476201102137566, + 0.004171865060925484, + 0.0026217785198241472, + 0.003089234931394458, + 0.004669813439249992, + 0.0023229201324284077, + 0.0025854867417365313, + 0.005179749336093664, + 0.003667650045827031, + 0.004119164310395718, + 0.002136440947651863, + 0.004247478675097227, + 0.0031570754945278168, + 0.002385364379733801, + 0.002833302365615964, + 0.003537863725796342, + 0.003322957782074809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.009806079789996147 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253913346678019 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253913346678019, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253913346678019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07944528013467789, + 0.07802789658308029, + 0.055556703358888626, + 0.49845969676971436, + 0.031059281900525093, + 0.07547592371702194, + 0.08607464283704758, + 0.050289787352085114, + 0.02661922387778759, + 0.01899155043065548 + ], + "credit_map": [ + 0.0073319380171597, + 0.009842201136052608, + 0.010449397377669811, + 0.008668567053973675, + 0.012689033523201942, + 0.01118107046931982, + 0.011267611756920815, + 0.008015462197363377, + 0.006713372189551592, + 0.010016262531280518, + 0.004987013526260853, + 0.010026018135249615, + 0.01150792557746172, + 0.010204694233834743, + 0.010312067344784737, + 0.014260541647672653, + 0.004958902020007372, + 0.0059705451130867004, + 0.004639904946088791, + 0.005617015063762665, + 0.006207186263054609, + 0.008468173444271088, + 0.01312474999576807, + 0.006570226047188044, + 0.02223479188978672, + 0.013450358994305134, + 0.016571221873164177, + 0.008073360659182072, + 0.1926768571138382, + 0.11231499165296555, + 0.02642657421529293, + 0.10671155154705048, + 0.002721119439229369, + 0.001645999145694077, + 0.003989986144006252, + 0.004784305114299059, + 0.003705363254994154, + 0.00470031937584281, + 0.00440774392336607, + 0.005104447714984417, + 0.008929700590670109, + 0.007968289777636528, + 0.008733670227229595, + 0.008356708101928234, + 0.011463179253041744, + 0.01243125181645155, + 0.009002097882330418, + 0.008591026067733765, + 0.009432544931769371, + 0.010546879842877388, + 0.008350582793354988, + 0.010182665660977364, + 0.010433287359774113, + 0.012058612890541553, + 0.015619504265487194, + 0.009450558573007584, + 0.006126635707914829, + 0.004753791727125645, + 0.0058301957324147224, + 0.007446084637194872, + 0.0060667311772704124, + 0.004703690763562918, + 0.01036195456981659, + 0.005000700708478689, + 0.0037761202547699213, + 0.00217240653000772, + 0.0032745692878961563, + 0.0037970428820699453, + 0.0026865683030337095, + 0.002747755032032728, + 0.005297735799103975, + 0.0028670283500105143, + 0.0027445314917713404, + 0.0014356609899550676, + 0.003022590419277549, + 0.002233614679425955, + 0.00316670723259449, + 0.0016278679249808192, + 0.002881661057472229, + 0.0018789158202707767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.005253913346678019 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985032498836517 + ], + "opposite_logprob_tokens": [ + -1.8048503398895264 + ], + "logprob_reward": -0.17985032498836517, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985032498836517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09928743541240692, + 0.09601057320833206, + 0.10137118399143219, + 0.3770488500595093, + 0.02346559800207615, + 0.055829692631959915, + 0.1618851125240326, + 0.058357495814561844, + 0.01688152551651001, + 0.009862583130598068 + ], + "credit_map": [ + 0.010318472050130367, + 0.010027483105659485, + 0.011339863762259483, + 0.007725196424871683, + 0.014533961191773415, + 0.014929182827472687, + 0.012259576469659805, + 0.01815369538962841, + 0.010949596762657166, + 0.009112919680774212, + 0.0092088608071208, + 0.011593452654778957, + 0.011276180855929852, + 0.011343706399202347, + 0.012454239651560783, + 0.020071618258953094, + 0.009312468580901623, + 0.0064909192733466625, + 0.009842791594564915, + 0.010517073795199394, + 0.012478603981435299, + 0.01604832150042057, + 0.024236757308244705, + 0.012444241903722286, + 0.01840919256210327, + 0.01367087010294199, + 0.01565926894545555, + 0.00882162619382143, + 0.13885489106178284, + 0.08014248311519623, + 0.022536052390933037, + 0.07895445823669434, + 0.0028061075136065483, + 0.0017834337195381522, + 0.002531511941924691, + 0.002714138710871339, + 0.003987215459346771, + 0.0035325787030160427, + 0.0025068274699151516, + 0.003603785065934062, + 0.005105083808302879, + 0.007035736460238695, + 0.00416178721934557, + 0.005872615613043308, + 0.006884305737912655, + 0.008465771563351154, + 0.006864245515316725, + 0.011440140195190907, + 0.01777324266731739, + 0.01981675624847412, + 0.017803506925702095, + 0.021023819223046303, + 0.0172884464263916, + 0.015437107533216476, + 0.03256230056285858, + 0.020179931074380875, + 0.007120692171156406, + 0.007304380182176828, + 0.006157257594168186, + 0.00824037566781044, + 0.007209676317870617, + 0.005690508056432009, + 0.010623172856867313, + 0.006011434830725193, + 0.002579264808446169, + 0.001442918204702437, + 0.0023692925460636616, + 0.0029646020848304033, + 0.0013557195197790861, + 0.001954055391252041, + 0.002779139671474695, + 0.0014365334063768387, + 0.0015362764243036509, + 0.000805414339993149, + 0.0011314358562231064, + 0.0007974366890266538, + 0.0020368071272969246, + 0.0009474293910898268, + 0.0019293386721983552, + 0.000678445037920028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.17985032498836517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, rendered in a gritty, unadorned photorealistic style, captures the raw power of a massive Hummer H1 Alpha as the camera begins a slow, steady pullback. Painted a matte olive drab green and almost entirely caked in thick, wet mud, the vehicle's iconic boxy shape laboriously climbs a steep, treacherous slope. Its massive, deeply treaded tires spin with immense torque, relentlessly digging into the saturated terrain and violently propelling arcs of dark brown soil and small rocks backward. The entire chassis exerts itself, the suspension articulating the noticeably uneven ground. The scene is bathed in the cool, unfocused light of an overcast sky, highlighting the glistening dampness of the tumbling mud and the splattered dirt clinging to the vehicle's chassis. A dense wall of dark green forest foliage, slightly out of focus in the background, combines with the dominant palette of dark browns and greens to create an atmosphere of intense, primal struggle between a powerful machine and the rugged landscape. The camera's slow, backward movement continuously emphasizes the vehicle's immense scale and the immense difficulty of the ascent.", + "seed": 694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_694_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": -0.04676549891009927, + "vlm_logprob": -0.04676549891009927, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04676549891009927, + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09689615666866302 + ], + "opposite_logprob_tokens": [ + -2.596896171569824 + ], + "logprob_reward": -0.09689615666866302, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09689615666866302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05526146665215492, + 0.03221079707145691, + 0.0324101522564888, + 0.03126048296689987, + 0.14937029778957367, + 0.03366352245211601, + 0.029690928757190704, + 0.0928487479686737, + 0.4631700813770294, + 0.08011341840028763 + ], + "credit_map": [ + 0.0062589519657194614, + 0.004552138969302177, + 0.0058065433986485004, + 0.008673533797264099, + 0.0067867799662053585, + 0.007500799372792244, + 0.010441646911203861, + 0.005241075064986944, + 0.0036692232824862003, + 0.003290684660896659, + 0.008380161598324776, + 0.003216077806428075, + 0.0024765925481915474, + 0.005161493085324764, + 0.0031390024814754725, + 0.002877563703805208, + 0.0027891856152564287, + 0.0036853919737040997, + 0.0039456370286643505, + 0.0047368621453642845, + 0.004250846803188324, + 0.004635389428585768, + 0.00503893755376339, + 0.0033279056660830975, + 0.001621350646018982, + 0.0038806083612143993, + 0.005718405824154615, + 0.004057807382196188, + 0.004065894056111574, + 0.0043308003805577755, + 0.004365039058029652, + 0.0032205800525844097, + 0.020353807136416435, + 0.01648334413766861, + 0.02522503025829792, + 0.017670422792434692, + 0.011721383780241013, + 0.0157251488417387, + 0.014681701548397541, + 0.027509475126862526, + 0.0037357998080551624, + 0.005205031484365463, + 0.005370069295167923, + 0.005838366225361824, + 0.0027284680400043726, + 0.003528980305418372, + 0.003775839926674962, + 0.003480971325188875, + 0.0022883396595716476, + 0.0035819069016724825, + 0.005015309434384108, + 0.003830722765997052, + 0.003668149933218956, + 0.0022697902750223875, + 0.005070071667432785, + 0.003966640215367079, + 0.006218092981725931, + 0.0052878158167004585, + 0.011636649258434772, + 0.02502826228737831, + 0.005035399459302425, + 0.0045624300837516785, + 0.013806707225739956, + 0.02127339318394661, + 0.11116986721754074, + 0.015610125847160816, + 0.04205687344074249, + 0.02962804213166237, + 0.017397195100784302, + 0.009801464155316353, + 0.19502270221710205, + 0.0424838587641716, + 0.005328114610165358, + 0.003677005646750331, + 0.012820389121770859, + 0.022565558552742004, + 0.004514614585787058, + 0.0038985637947916985, + 0.016662154346704483, + 0.010647022165358067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.09689615666866302 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005307393614202738 + ], + "opposite_logprob_tokens": [ + -5.255307197570801 + ], + "logprob_reward": -0.005307393614202738, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005307393614202738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21476590633392334, + 0.12370054423809052, + 0.03769639506936073, + 0.05025901272892952, + 0.13770677149295807, + 0.03883948549628258, + 0.026654506102204323, + 0.14707051217556, + 0.10612620413303375, + 0.1171807125210762 + ], + "credit_map": [ + 0.03193852677941322, + 0.013600043021142483, + 0.02552250400185585, + 0.031585924327373505, + 0.023692233487963676, + 0.03146515041589737, + 0.034210942685604095, + 0.02275056578218937, + 0.01927116885781288, + 0.013295795768499374, + 0.025015993043780327, + 0.013126431033015251, + 0.01171770691871643, + 0.01978665217757225, + 0.011147215962409973, + 0.010339575819671154, + 0.002786917146295309, + 0.004840158391743898, + 0.005397239699959755, + 0.005946025252342224, + 0.00459875026717782, + 0.005310057196766138, + 0.005196465644985437, + 0.003620777977630496, + 0.004579131957143545, + 0.004780183546245098, + 0.011949478648602962, + 0.0041318065486848354, + 0.004804720636457205, + 0.00835823081433773, + 0.005879240110516548, + 0.005776220466941595, + 0.0188093651086092, + 0.015561879612505436, + 0.02421010099351406, + 0.017059609293937683, + 0.010343078523874283, + 0.013655644841492176, + 0.013299885205924511, + 0.02476719208061695, + 0.004110265523195267, + 0.004711936227977276, + 0.007216206286102533, + 0.009091409854590893, + 0.002264787908643484, + 0.0036749111022800207, + 0.0040413569658994675, + 0.0037286097649484873, + 0.0029292397666722536, + 0.003026018850505352, + 0.003838552162051201, + 0.0031379505526274443, + 0.0031055710278451443, + 0.0022423844784498215, + 0.0035443406086415052, + 0.004830446094274521, + 0.009470818564295769, + 0.010327063500881195, + 0.02327464148402214, + 0.05879686027765274, + 0.005455465521663427, + 0.004129005596041679, + 0.02143566496670246, + 0.014180988073348999, + 0.022156408056616783, + 0.008236579596996307, + 0.010981493629515171, + 0.00961310788989067, + 0.005934778600931168, + 0.003683379851281643, + 0.034445781260728836, + 0.011074669659137726, + 0.00899828877300024, + 0.009986670687794685, + 0.03145061433315277, + 0.028291376307606697, + 0.006538974586874247, + 0.005049353465437889, + 0.013486240059137344, + 0.013379178941249847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.005307393614202738 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091247200965881 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091247200965881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091247200965881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11265473067760468, + 0.05721784383058548, + 0.04672649875283241, + 0.04344286769628525, + 0.3203390836715698, + 0.07046651840209961, + 0.03686210513114929, + 0.08947832882404327, + 0.18003962934017181, + 0.04277229309082031 + ], + "credit_map": [ + 0.013741453178226948, + 0.00697639724239707, + 0.01886618882417679, + 0.01116232667118311, + 0.011361992917954922, + 0.01638181321322918, + 0.02193126082420349, + 0.012233312241733074, + 0.004802864044904709, + 0.004228736739605665, + 0.009361782111227512, + 0.008488647639751434, + 0.00801151618361473, + 0.007726417854428291, + 0.007871054112911224, + 0.006726831663399935, + 0.0034390182700008154, + 0.005738409236073494, + 0.006226667668670416, + 0.007393901702016592, + 0.004777226597070694, + 0.006931223440915346, + 0.008303782902657986, + 0.003916275221854448, + 0.004761474207043648, + 0.00544347520917654, + 0.00893673300743103, + 0.004131382796913385, + 0.005545893218368292, + 0.005923576653003693, + 0.004708097781985998, + 0.003992235753685236, + 0.03998028486967087, + 0.03504526987671852, + 0.05419136956334114, + 0.03773434832692146, + 0.026795431971549988, + 0.03289485350251198, + 0.03447327762842178, + 0.05922427773475647, + 0.012601366266608238, + 0.008091757073998451, + 0.011620533652603626, + 0.014807652682065964, + 0.004680547397583723, + 0.005742151290178299, + 0.006226828787475824, + 0.006695690099149942, + 0.003409744007512927, + 0.0036300078500062227, + 0.004571116529405117, + 0.006787516176700592, + 0.007354805711656809, + 0.0034832630772143602, + 0.004983790684491396, + 0.0026418622583150864, + 0.007478374522179365, + 0.005554911680519581, + 0.014376464299857616, + 0.02919076569378376, + 0.007263341918587685, + 0.004862030502408743, + 0.01453468855470419, + 0.006217753980308771, + 0.03716582804918289, + 0.009431581012904644, + 0.019126422703266144, + 0.012827829457819462, + 0.007832219824194908, + 0.005286088678985834, + 0.07091067731380463, + 0.01745898835361004, + 0.0037298572715371847, + 0.004061813931912184, + 0.007826762273907661, + 0.005995063576847315, + 0.003369763493537903, + 0.0028610853478312492, + 0.009049355052411556, + 0.005878595635294914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.011091247200965881 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076294224709272385 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.0076294224709272385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076294224709272385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09068910032510757, + 0.10038738697767258, + 0.03342708200216293, + 0.03593847155570984, + 0.2540055513381958, + 0.051087018102407455, + 0.05410650372505188, + 0.1465943157672882, + 0.12535053491592407, + 0.10841400176286697 + ], + "credit_map": [ + 0.011426305398344994, + 0.008587021380662918, + 0.01544973999261856, + 0.008805508725345135, + 0.010674748569726944, + 0.01273458730429411, + 0.013224206864833832, + 0.009786981157958508, + 0.009240927174687386, + 0.011762306094169617, + 0.02326383627951145, + 0.01295387465506792, + 0.007919383235275745, + 0.01570010371506214, + 0.010280648246407509, + 0.009266314096748829, + 0.0023522176779806614, + 0.0043890271335840225, + 0.004729233216494322, + 0.004435027949512005, + 0.0047352127730846405, + 0.004514059517532587, + 0.005211202893406153, + 0.003061098512262106, + 0.001972212688997388, + 0.0036679033655673265, + 0.0069137257523834705, + 0.004875284619629383, + 0.002507540863007307, + 0.00475023640319705, + 0.0060431151650846004, + 0.0052084545604884624, + 0.03410657122731209, + 0.02862272411584854, + 0.04390779137611389, + 0.03131324425339699, + 0.020810507237911224, + 0.025426678359508514, + 0.024837521836161613, + 0.04498053714632988, + 0.006067126523703337, + 0.006041146349161863, + 0.008102063089609146, + 0.011796063743531704, + 0.004338970873504877, + 0.004025558475404978, + 0.005839820485562086, + 0.004876269958913326, + 0.006213524378836155, + 0.005394979380071163, + 0.006765122525393963, + 0.00895668100565672, + 0.00827792752534151, + 0.004150150343775749, + 0.008159087039530277, + 0.006189032457768917, + 0.01181539986282587, + 0.007453999947756529, + 0.029193194583058357, + 0.026665640994906425, + 0.008872938342392445, + 0.007780609652400017, + 0.021524833515286446, + 0.033287689089775085, + 0.02761111594736576, + 0.004301765002310276, + 0.012680547311902046, + 0.008439486846327782, + 0.005742660257965326, + 0.0027832111809402704, + 0.051999568939208984, + 0.011792161501944065, + 0.007877785712480545, + 0.006559750530868769, + 0.01605895161628723, + 0.03961118683218956, + 0.003688315860927105, + 0.003926737233996391, + 0.016238242387771606, + 0.01445303950458765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0076294224709272385 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290327459573746 + ], + "opposite_logprob_tokens": [ + -2.237903356552124 + ], + "logprob_reward": -0.11290327459573746, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290327459573746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037704091519117355, + 0.03310096263885498, + 0.022147048264741898, + 0.02666676975786686, + 0.17758657038211823, + 0.024161307141184807, + 0.044696249067783356, + 0.10194305330514908, + 0.498393177986145, + 0.0336008258163929 + ], + "credit_map": [ + 0.0036480252165347338, + 0.0025602378882467747, + 0.006752669345587492, + 0.004738607909530401, + 0.006185656413435936, + 0.0061315447092056274, + 0.005225399509072304, + 0.0024619472678750753, + 0.0026035066694021225, + 0.0034424394834786654, + 0.005023420788347721, + 0.004319879692047834, + 0.004197344649583101, + 0.004525368567556143, + 0.005861206445842981, + 0.0031277958769351244, + 0.0023761424235999584, + 0.002403191989287734, + 0.0029651978984475136, + 0.0029670309741050005, + 0.003167799673974514, + 0.002690924098715186, + 0.0034351320937275887, + 0.0021416270174086094, + 0.0029196864925324917, + 0.002758044982329011, + 0.0048367902636528015, + 0.0033769942820072174, + 0.0024142267648130655, + 0.003491699229925871, + 0.003597230650484562, + 0.003272094763815403, + 0.02392815239727497, + 0.019457340240478516, + 0.030383437871932983, + 0.02194913476705551, + 0.01405579224228859, + 0.01772274635732174, + 0.01820436306297779, + 0.03188559040427208, + 0.00218877661973238, + 0.003155392361804843, + 0.0032893556635826826, + 0.004412357229739428, + 0.0022965618409216404, + 0.002286059083417058, + 0.0036518233828246593, + 0.0028809804935008287, + 0.005608315579593182, + 0.00442240247502923, + 0.006101047620177269, + 0.006496482063084841, + 0.006505491677671671, + 0.0032505991403013468, + 0.006392138544470072, + 0.005919768940657377, + 0.009699284099042416, + 0.004862596280872822, + 0.027536358684301376, + 0.010243911296129227, + 0.005186562426388264, + 0.005300070624798536, + 0.014785693027079105, + 0.02432856895029545, + 0.11948874592781067, + 0.01693340390920639, + 0.045277249068021774, + 0.03191148117184639, + 0.01938377134501934, + 0.010460239835083485, + 0.2090238332748413, + 0.045914426445961, + 0.002251456258818507, + 0.0015459548449143767, + 0.008782932534813881, + 0.012059997767210007, + 0.0013167430879548192, + 0.0010451977141201496, + 0.0037445419002324343, + 0.002853997051715851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.11290327459573746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, rendered in a gritty, unadorned photorealistic style, captures the raw power of a massive Hummer H1 Alpha as the camera begins a slow, steady pullback. Painted a matte olive drab green and almost entirely caked in thick, wet mud, the vehicle's iconic boxy shape laboriously climbs a steep, treacherous slope. Its massive, deeply treaded tires spin with immense torque, relentlessly digging into the saturated terrain and violently propelling arcs of dark brown soil and small rocks backward. The entire chassis exerts itself, the suspension articulating the noticeably uneven ground. The scene is bathed in the cool, unfocused light of an overcast sky, highlighting the glistening dampness of the tumbling mud and the splattered dirt clinging to the vehicle's chassis. A dense wall of dark green forest foliage, slightly out of focus in the background, combines with the dominant palette of dark browns and greens to create an atmosphere of intense, primal struggle between a powerful machine and the rugged landscape. The camera's slow, backward movement continuously emphasizes the vehicle's immense scale and the immense difficulty of the ascent.", + "seed": 695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_13/index_173_seed_695_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": -0.03550144918262958, + "vlm_logprob": -0.03550144918262958, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.03550144918262958, + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08404028415679932 + ], + "opposite_logprob_tokens": [ + -2.7090401649475098 + ], + "logprob_reward": -0.08404028415679932, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08404028415679932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10302506387233734, + 0.5340301990509033, + 0.05224990472197533, + 0.0334295816719532, + 0.14298927783966064, + 0.02181318588554859, + 0.021552227437496185, + 0.04228464886546135, + 0.030335450544953346, + 0.018290510401129723 + ], + "credit_map": [ + 0.01110732089728117, + 0.00938370916992426, + 0.013880254700779915, + 0.012618528679013252, + 0.013589262031018734, + 0.018584439530968666, + 0.011004261672496796, + 0.01285728719085455, + 0.09571117907762527, + 0.025205569341778755, + 0.030991001054644585, + 0.02912706881761551, + 0.024249577894806862, + 0.24547629058361053, + 0.04660411924123764, + 0.0366654098033905, + 0.004110518377274275, + 0.0029852830339223146, + 0.002563654677942395, + 0.004044543486088514, + 0.0032849465496838093, + 0.01713588275015354, + 0.004042861517518759, + 0.014082214795053005, + 0.0037321713753044605, + 0.005042841657996178, + 0.0032115194480866194, + 0.005644385237246752, + 0.0026158392429351807, + 0.0025837370194494724, + 0.005824373569339514, + 0.004774713423103094, + 0.013107344508171082, + 0.013886190950870514, + 0.023896614089608192, + 0.019545866176486015, + 0.012270539999008179, + 0.009697197005152702, + 0.031205805018544197, + 0.019379721954464912, + 0.0024653994478285313, + 0.0029161216225475073, + 0.0025324702728539705, + 0.002806637668982148, + 0.0024445480667054653, + 0.003152970690280199, + 0.0027710788417607546, + 0.0027239606715738773, + 0.0032319393940269947, + 0.003256311872974038, + 0.0023841748479753733, + 0.003038351656869054, + 0.00199874653480947, + 0.0023023064713925123, + 0.002597876824438572, + 0.002742517041042447, + 0.004085128661245108, + 0.007647685706615448, + 0.004897051025182009, + 0.003729404415935278, + 0.005700206849724054, + 0.004768745508044958, + 0.006455122493207455, + 0.005001308396458626, + 0.004405740182846785, + 0.0025149902794510126, + 0.004044996574521065, + 0.008625519461929798, + 0.0015168952522799373, + 0.0021022844593971968, + 0.003502750303596258, + 0.0036222708877176046, + 0.0026110978797078133, + 0.0030754057224839926, + 0.0020589467603713274, + 0.0020647093188017607, + 0.002230057492852211, + 0.0025952993892133236, + 0.0018918211571872234, + 0.0017631720984354615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.08404028415679932 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036662286147475243 + ], + "opposite_logprob_tokens": [ + -5.628666400909424 + ], + "logprob_reward": -0.0036662286147475243, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036662286147475243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16840843856334686, + 0.189261794090271, + 0.1186814233660698, + 0.04023490101099014, + 0.09972846508026123, + 0.0606512725353241, + 0.03572510555386543, + 0.075987309217453, + 0.1743619292974472, + 0.03695933148264885 + ], + "credit_map": [ + 0.024231579154729843, + 0.01714637503027916, + 0.025306353345513344, + 0.020306071266531944, + 0.027123764157295227, + 0.01824183017015457, + 0.015894567593932152, + 0.020157907158136368, + 0.028867879882454872, + 0.015447951853275299, + 0.013718098402023315, + 0.01847800612449646, + 0.023478996008634567, + 0.05939844995737076, + 0.017446812242269516, + 0.012425603345036507, + 0.010109488852322102, + 0.006227897945791483, + 0.00434507941827178, + 0.008067266084253788, + 0.006694297306239605, + 0.04117336869239807, + 0.008665462955832481, + 0.03339856490492821, + 0.004459375515580177, + 0.006574881263077259, + 0.003624392906203866, + 0.006580385845154524, + 0.0035323926713317633, + 0.004843549337238073, + 0.005653491243720055, + 0.004966434091329575, + 0.009364164434373379, + 0.009924652986228466, + 0.01959056407213211, + 0.011829386465251446, + 0.008587492629885674, + 0.0061790067702531815, + 0.02054673619568348, + 0.013706456869840622, + 0.004496438428759575, + 0.013689461164176464, + 0.005866909399628639, + 0.016389617696404457, + 0.002748288447037339, + 0.0027723731473088264, + 0.008155071176588535, + 0.0065331184305250645, + 0.003553746035322547, + 0.004745286423712969, + 0.0036763669922947884, + 0.00464906869456172, + 0.0029661699663847685, + 0.008371660485863686, + 0.004128037486225367, + 0.0036347676068544388, + 0.006909196265041828, + 0.009441210888326168, + 0.01058997306972742, + 0.006236208137124777, + 0.007103417534381151, + 0.011834761127829552, + 0.014265288598835468, + 0.009607253596186638, + 0.02461129054427147, + 0.01626497134566307, + 0.020447857677936554, + 0.029912354424595833, + 0.010725966654717922, + 0.01594356819987297, + 0.017333699390292168, + 0.03912222012877464, + 0.006193162407726049, + 0.0038990278262645006, + 0.004958125296980143, + 0.005555161274969578, + 0.0031081773340702057, + 0.004216492176055908, + 0.004740006756037474, + 0.004289175849407911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0036662286147475243 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111053422093391 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111053422093391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111053422093391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24418681859970093, + 0.1357073038816452, + 0.2689509987831116, + 0.0322163887321949, + 0.10129831731319427, + 0.044150516390800476, + 0.04465892165899277, + 0.022063158452510834, + 0.0758972018957138, + 0.030870379880070686 + ], + "credit_map": [ + 0.030255554243922234, + 0.02437320165336132, + 0.04548444226384163, + 0.025713738054037094, + 0.0322980172932148, + 0.028063025325536728, + 0.017187654972076416, + 0.04081117734313011, + 0.019789118319749832, + 0.009151509031653404, + 0.009351447224617004, + 0.00894884206354618, + 0.013698420487344265, + 0.051769644021987915, + 0.013362931087613106, + 0.00963540654629469, + 0.016545947641134262, + 0.008703173138201237, + 0.0051368060521781445, + 0.01454667467623949, + 0.00912314373999834, + 0.10114920884370804, + 0.014029895886778831, + 0.09971616417169571, + 0.0035943782422691584, + 0.0036555128172039986, + 0.004929214250296354, + 0.0040871077217161655, + 0.0025104391388595104, + 0.005402090027928352, + 0.004118508193641901, + 0.003919138573110104, + 0.00980300921946764, + 0.010244985111057758, + 0.021778525784611702, + 0.015487780794501305, + 0.006490237079560757, + 0.007780290674418211, + 0.013462920673191547, + 0.01625056564807892, + 0.0036204447969794273, + 0.01141244824975729, + 0.0030834609642624855, + 0.012013129889965057, + 0.002183311153203249, + 0.0026250670198351145, + 0.004949701949954033, + 0.004262950737029314, + 0.004063554108142853, + 0.005600780714303255, + 0.006455949507653713, + 0.006180791649967432, + 0.0038756851572543383, + 0.005823036655783653, + 0.007451596669852734, + 0.005207529291510582, + 0.002366434084251523, + 0.0026411847211420536, + 0.004900303203612566, + 0.0020840049255639315, + 0.002439577830955386, + 0.0022294309455901384, + 0.0027316745836287737, + 0.002670548390597105, + 0.010351261124014854, + 0.006533428095281124, + 0.008160384371876717, + 0.014556588605046272, + 0.0047418056055903435, + 0.008389292284846306, + 0.008782722055912018, + 0.014381720684468746, + 0.0053300210274755955, + 0.002753720385953784, + 0.005195865407586098, + 0.004658011253923178, + 0.0026462827809154987, + 0.003617340698838234, + 0.003692406229674816, + 0.002976731862872839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004111053422093391 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737136282026768 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737136282026768, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737136282026768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10149713605642319, + 0.3172815442085266, + 0.0414353609085083, + 0.023908253759145737, + 0.07313987612724304, + 0.03440936654806137, + 0.019481347873806953, + 0.02912001870572567, + 0.32364508509635925, + 0.036082036793231964 + ], + "credit_map": [ + 0.01724417880177498, + 0.008316085673868656, + 0.01682593673467636, + 0.010700198821723461, + 0.010540959425270557, + 0.014867733232676983, + 0.0089354757219553, + 0.014066558331251144, + 0.05649198591709137, + 0.012408983893692493, + 0.018499502912163734, + 0.015738418325781822, + 0.014585846103727818, + 0.14922280609607697, + 0.02660280652344227, + 0.023731151595711708, + 0.00395806971937418, + 0.0023931784089654684, + 0.00192918349057436, + 0.0032162750139832497, + 0.004433714784681797, + 0.007864611223340034, + 0.0033283131197094917, + 0.014312009327113628, + 0.0026008528657257557, + 0.00333007681183517, + 0.003667730139568448, + 0.003401488997042179, + 0.0018160276813432574, + 0.0018807149026542902, + 0.004868319723755121, + 0.0023430390283465385, + 0.007490042597055435, + 0.0057898107916116714, + 0.011048628017306328, + 0.010289848782122135, + 0.004978318233042955, + 0.006175082176923752, + 0.014135654084384441, + 0.013232480734586716, + 0.0035136365331709385, + 0.006619628984481096, + 0.0031449380330741405, + 0.005101102404296398, + 0.0023568090982735157, + 0.002686308929696679, + 0.005633367691189051, + 0.0053535704500973225, + 0.001966390060260892, + 0.0017477998044341803, + 0.004194270353764296, + 0.002746002981439233, + 0.0015277772909030318, + 0.0030308347195386887, + 0.002642529783770442, + 0.0016257403185591102, + 0.003176054684445262, + 0.003918417263776064, + 0.006075022742152214, + 0.002751904306933284, + 0.002460482530295849, + 0.0027889308985322714, + 0.00502369087189436, + 0.0029255137778818607, + 0.045491188764572144, + 0.02993123047053814, + 0.034553419798612595, + 0.05725305527448654, + 0.02003665640950203, + 0.031008441001176834, + 0.03307218849658966, + 0.07229886949062347, + 0.00534528773277998, + 0.0037189589347690344, + 0.005090230144560337, + 0.005867678672075272, + 0.0029054651968181133, + 0.005228809081017971, + 0.0041047376580536366, + 0.0038208647165447474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.006737136282026768 + }, + { + "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 Hummer H1 Alpha painted matte olive drab green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the vehicle almost entirely covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vehicle's deeply treaded tires spinning and digging into the terrain, propelling arcs of dark brown soil and rocks backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a steep, treacherous slope with a dense wall of dark green forest foliage in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and execute a slow, steady pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895254343748093 + ], + "opposite_logprob_tokens": [ + -2.5789525508880615 + ], + "logprob_reward": -0.07895254343748093, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895254343748093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11254896223545074, + 0.21661953628063202, + 0.07953926175832748, + 0.05572880804538727, + 0.131553515791893, + 0.16920965909957886, + 0.10743159055709839, + 0.04790382459759712, + 0.04181838035583496, + 0.037646520882844925 + ], + "credit_map": [ + 0.01579066552221775, + 0.009713219478726387, + 0.014829565770924091, + 0.009544105269014835, + 0.015385852195322514, + 0.01815708912909031, + 0.01134261954575777, + 0.01778583414852619, + 0.03388543426990509, + 0.011710991151630878, + 0.013045044615864754, + 0.011874649673700333, + 0.012268474325537682, + 0.09255542606115341, + 0.02035418711602688, + 0.020925307646393776, + 0.008035511709749699, + 0.00884860660880804, + 0.007820913568139076, + 0.011598595418035984, + 0.010240734554827213, + 0.01065407507121563, + 0.01319872122257948, + 0.009142087772488594, + 0.0051226238720119, + 0.006757600232958794, + 0.011419889517128468, + 0.0072278818115592, + 0.004288094583898783, + 0.004572044126689434, + 0.008419863879680634, + 0.007920805364847183, + 0.014179818332195282, + 0.012139943428337574, + 0.03912385180592537, + 0.01746581494808197, + 0.010537481866776943, + 0.007145251147449017, + 0.017984503880143166, + 0.012976834550499916, + 0.015462329611182213, + 0.04878043383359909, + 0.012049072422087193, + 0.04392723739147186, + 0.008046205155551434, + 0.009490328840911388, + 0.01609884388744831, + 0.015355191193521023, + 0.0099002905189991, + 0.010873273946344852, + 0.01858457922935486, + 0.01707202009856701, + 0.007675654254853725, + 0.020524706691503525, + 0.01300985086709261, + 0.009791207499802113, + 0.005174565128982067, + 0.005566163454204798, + 0.00842069461941719, + 0.00556725449860096, + 0.00421637250110507, + 0.005409130826592445, + 0.008391903713345528, + 0.00515773706138134, + 0.0064885844476521015, + 0.0047313193790614605, + 0.005265343002974987, + 0.005394004750996828, + 0.0033508080523461103, + 0.004282496869564056, + 0.004909292329102755, + 0.007396525703370571, + 0.007265880703926086, + 0.004355061333626509, + 0.004427962005138397, + 0.007559024263173342, + 0.002965005347505212, + 0.0032557747326791286, + 0.003391029080376029, + 0.004426778759807348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.07895254343748093 + } + ], + "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/0000017/group_14/index_tensor([77, 77])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_tensor([77, 77])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ac6898c3f06b7c51c213687c5ebfa3ae67f46056 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_tensor([77, 77])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera executes a dramatic, low-angle, spiraling orbital shot, circling an upward flight of gleaming metallic robots as they blast upward in perfect unison from a circular, underground launch pad. Each robot is constructed of polished gunmetal-gray alloy to a mirror-like sheen, with a sleek, aerodynamic humanoid shape and integrated flight stabilizers on its back and limbs. Their featureless obsidian visors glow with an intense inner white light, and thin cobalt-blue energy conduits trace the seams of their armor plating. The action begins as they simultaneously ignite the powerful plasma thrusters beneath their feet. The initial blast of fiery blue energy shatters the concrete floor beneath them, sending a torrent of steam and debris cascading outward. As they ascend at breathtaking speed into a vast, dark silo, the camera stays in sync with them, perfectly executing a series of complex flips—a tight barrel roll before rapidly dispersing and regrouping into a precise diamond formation. The only sources of light are the robots’ own bright thruster plumes and glowing components, which cast harsh, rapidly changing highlights and deep, dancing shadows across their metallic bodies, creating an atmosphere of high contrast, intense dynamism, and breathtaking, overwhelming technological power.", + "seed": 12991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_77_seed_12991_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": -0.18315375670790673, + "vlm_logprob": -0.18315375670790673, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.18315375670790673, + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09483064711093903 + ], + "opposite_logprob_tokens": [ + -2.5948307514190674 + ], + "logprob_reward": -0.09483064711093903, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09483064711093903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08995667845010757, + 0.1184011772274971, + 0.07083112001419067, + 0.03331470489501953, + 0.03285897150635719, + 0.07116961479187012, + 0.10083187371492386, + 0.23385600745677948, + 0.1841331124305725, + 0.06464681029319763 + ], + "credit_map": [ + 0.011209134943783283, + 0.009034084156155586, + 0.006984067615121603, + 0.008412507362663746, + 0.008568362332880497, + 0.013572528026998043, + 0.022530237212777138, + 0.009645753540098667, + 0.01555780041962862, + 0.02072865702211857, + 0.01395410392433405, + 0.017384052276611328, + 0.013731031678617, + 0.015708228573203087, + 0.009079056791961193, + 0.01225823163986206, + 0.009984874166548252, + 0.005844017490744591, + 0.0037148126866668463, + 0.004280510824173689, + 0.004493783228099346, + 0.019646408036351204, + 0.002455794718116522, + 0.02041091024875641, + 0.003010724438354373, + 0.005414496175944805, + 0.003520892234519124, + 0.004478771705180407, + 0.0060075693763792515, + 0.00458300206810236, + 0.0031089310068637133, + 0.0031903181225061417, + 0.005497651174664497, + 0.003075107466429472, + 0.00371807673946023, + 0.0042752549052238464, + 0.004605315625667572, + 0.0036659520119428635, + 0.0035409750416874886, + 0.004480637144297361, + 0.011523718945682049, + 0.009919078089296818, + 0.0066298190504312515, + 0.008167929947376251, + 0.012769930996000767, + 0.009732300415635109, + 0.005097287241369486, + 0.007329539395868778, + 0.01547622587531805, + 0.014080571010708809, + 0.00997285544872284, + 0.010706299915909767, + 0.010979178361594677, + 0.011167573742568493, + 0.010130498558282852, + 0.0183186586946249, + 0.0712784081697464, + 0.007843784056603909, + 0.012869811616837978, + 0.007941285148262978, + 0.009779890067875385, + 0.08835240453481674, + 0.009576290845870972, + 0.026214128360152245, + 0.021173132583498955, + 0.024531207978725433, + 0.02273644134402275, + 0.028132345527410507, + 0.024003252387046814, + 0.02959752269089222, + 0.01700693741440773, + 0.016952261328697205, + 0.019651679322123528, + 0.0035991999320685863, + 0.016679784283041954, + 0.004077258985489607, + 0.003307695733383298, + 0.00846215058118105, + 0.0032602648716419935, + 0.005608772858977318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.09483064711093903 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636663056910038 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.02636663056910038, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636663056910038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0409870482981205, + 0.040653735399246216, + 0.1996193528175354, + 0.02443026192486286, + 0.03531268611550331, + 0.052288398146629333, + 0.15097832679748535, + 0.3304966688156128, + 0.05108361691236496, + 0.07414986938238144 + ], + "credit_map": [ + 0.005580407567322254, + 0.0040106577798724174, + 0.003914198372513056, + 0.005153245758265257, + 0.003742886707186699, + 0.005660390947014093, + 0.008236294612288475, + 0.004688967019319534, + 0.0052379099652171135, + 0.0067367879673838615, + 0.003765533212572336, + 0.005314880050718784, + 0.005743875168263912, + 0.005123992916196585, + 0.0029892977327108383, + 0.0057414560578763485, + 0.009457051753997803, + 0.015544235706329346, + 0.0059313648380339146, + 0.004928251262754202, + 0.0043755462393164635, + 0.08120503276586533, + 0.003421319415792823, + 0.07475654780864716, + 0.00275484099984169, + 0.005422513000667095, + 0.0017289402894675732, + 0.001842502155341208, + 0.005357668269425631, + 0.0029706666246056557, + 0.0021186573430895805, + 0.0022344724275171757, + 0.005300108343362808, + 0.0035961796529591084, + 0.0038968136068433523, + 0.0025706500746309757, + 0.005480635911226273, + 0.004486371297389269, + 0.002440983196720481, + 0.0075409444980323315, + 0.009710337035357952, + 0.006752359680831432, + 0.007467504590749741, + 0.005432046018540859, + 0.007309313397854567, + 0.006670474074780941, + 0.004076366312801838, + 0.004869992844760418, + 0.01759527251124382, + 0.02127024717628956, + 0.020551534369587898, + 0.012196683324873447, + 0.025357205420732498, + 0.01950613036751747, + 0.014294786378741264, + 0.02020646072924137, + 0.10078033059835434, + 0.011640748009085655, + 0.01826145499944687, + 0.010176904499530792, + 0.012879536487162113, + 0.12783636152744293, + 0.012622041627764702, + 0.036299269646406174, + 0.0070344228297472, + 0.01026792824268341, + 0.0048411632888019085, + 0.00433684978634119, + 0.00814010202884674, + 0.006116232369095087, + 0.0054047005251049995, + 0.004942222032696009, + 0.023993153125047684, + 0.0039151436649262905, + 0.020714301615953445, + 0.0036065801978111267, + 0.0030099658761173487, + 0.009378191083669662, + 0.002590300515294075, + 0.006942234933376312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.02636663056910038 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091125085949898 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091125085949898, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091125085949898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19241052865982056, + 0.07311694324016571, + 0.18380463123321533, + 0.03187200054526329, + 0.03326081857085228, + 0.04253946244716644, + 0.11011845618486404, + 0.16604460775852203, + 0.09101922810077667, + 0.07581333070993423 + ], + "credit_map": [ + 0.01954239420592785, + 0.021144306287169456, + 0.02244529314339161, + 0.024963458999991417, + 0.01718822866678238, + 0.025856902822852135, + 0.03992156684398651, + 0.021348362788558006, + 0.008468974381685257, + 0.0063971723429858685, + 0.008335509337484837, + 0.01292412169277668, + 0.008365758694708347, + 0.010266559198498726, + 0.0053616235964000225, + 0.012997228652238846, + 0.009927169419825077, + 0.017791152000427246, + 0.005993410479277372, + 0.007089470978826284, + 0.005168406292796135, + 0.07103481888771057, + 0.004708330612629652, + 0.06209186464548111, + 0.0030298209749162197, + 0.00552747305482626, + 0.0026307690422981977, + 0.0037056016735732555, + 0.00421086186543107, + 0.004301097709685564, + 0.005212268326431513, + 0.003254109527915716, + 0.004903580993413925, + 0.005464893765747547, + 0.0036007878370583057, + 0.0035727936774492264, + 0.005562069360166788, + 0.003922991920262575, + 0.0023636461701244116, + 0.0038700555451214314, + 0.006200452335178852, + 0.006349906325340271, + 0.005190332420170307, + 0.005844572558999062, + 0.004900822881609201, + 0.005756160710006952, + 0.004346077796071768, + 0.003951136488467455, + 0.019766179844737053, + 0.010557260364294052, + 0.012688903138041496, + 0.01032155379652977, + 0.011417282745242119, + 0.011342383921146393, + 0.010752982459962368, + 0.023271916434168816, + 0.04950496554374695, + 0.009788882918655872, + 0.009798943065106869, + 0.007318056188523769, + 0.009022969752550125, + 0.05680183321237564, + 0.006649283226579428, + 0.01715967431664467, + 0.017551153898239136, + 0.009457984939217567, + 0.008490679785609245, + 0.009512928314507008, + 0.009368415921926498, + 0.01280582882463932, + 0.006721076089888811, + 0.01711116172373295, + 0.021044986322522163, + 0.004531515296548605, + 0.020237816497683525, + 0.0037853519897907972, + 0.005078129004687071, + 0.008643999695777893, + 0.0033251929562538862, + 0.009166333824396133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.016091125085949898 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588964223861694 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588964223861694, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588964223861694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08832941204309464, + 0.20245996117591858, + 0.057644590735435486, + 0.06146276742219925, + 0.05134640261530876, + 0.12096022814512253, + 0.11001758277416229, + 0.07782687246799469, + 0.17520761489868164, + 0.054744526743888855 + ], + "credit_map": [ + 0.009449590928852558, + 0.007447327487170696, + 0.008891069330275059, + 0.009602934122085571, + 0.0066807870753109455, + 0.01381719671189785, + 0.015376070514321327, + 0.01706443727016449, + 0.025367630645632744, + 0.01876792497932911, + 0.026312094181776047, + 0.02441166341304779, + 0.029837267473340034, + 0.027294667437672615, + 0.01963837631046772, + 0.030830340459942818, + 0.009786579757928848, + 0.007755801547318697, + 0.0053609334863722324, + 0.004624225664883852, + 0.005350680556148291, + 0.01224531140178442, + 0.002855895785614848, + 0.009665163233876228, + 0.007013946305960417, + 0.005976978223770857, + 0.007136298343539238, + 0.009650377556681633, + 0.009026644751429558, + 0.005768570117652416, + 0.009704421274363995, + 0.0071855285204946995, + 0.007816941477358341, + 0.006688878871500492, + 0.004818783141672611, + 0.005666065029799938, + 0.006612242199480534, + 0.006907029077410698, + 0.005945216864347458, + 0.006891245488077402, + 0.014698902145028114, + 0.013270986266434193, + 0.014842020347714424, + 0.02456749603152275, + 0.012586157768964767, + 0.014483662322163582, + 0.017697861418128014, + 0.008813144639134407, + 0.015315464697778225, + 0.010586321353912354, + 0.017106615006923676, + 0.008802109397947788, + 0.015095041133463383, + 0.018704919144511223, + 0.009256563149392605, + 0.015150545164942741, + 0.023112913593649864, + 0.004446262493729591, + 0.004987143445760012, + 0.004158844240009785, + 0.00446319580078125, + 0.026096586138010025, + 0.003744371235370636, + 0.006817560642957687, + 0.032300252467393875, + 0.019970117136836052, + 0.017862148582935333, + 0.016757143661379814, + 0.018459396436810493, + 0.019507231190800667, + 0.017336536198854446, + 0.033014774322509766, + 0.006656335201114416, + 0.0067866346798837185, + 0.005097236018627882, + 0.006057150661945343, + 0.008483137004077435, + 0.008443872444331646, + 0.005852671805769205, + 0.007367489393800497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.020588964223861694 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578914165496826 + ], + "opposite_logprob_tokens": [ + -0.6328914165496826 + ], + "logprob_reward": -0.7578914165496826, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578914165496826, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13638287782669067, + 0.08493135869503021, + 0.20083141326904297, + 0.03794960677623749, + 0.0724078044295311, + 0.11559326201677322, + 0.10648870468139648, + 0.07283250242471695, + 0.11186879128217697, + 0.06071365997195244 + ], + "credit_map": [ + 0.018769264221191406, + 0.016137510538101196, + 0.018137535080313683, + 0.022742072120308876, + 0.01054976973682642, + 0.0121064567938447, + 0.015136691741645336, + 0.022803574800491333, + 0.008743179962038994, + 0.01290466170758009, + 0.007490110583603382, + 0.011746290139853954, + 0.0113975303247571, + 0.0113243218511343, + 0.006941396743059158, + 0.01438386831432581, + 0.010416531004011631, + 0.015067776665091515, + 0.0054413690231740475, + 0.00777546688914299, + 0.004756812006235123, + 0.07721357047557831, + 0.0045762211084365845, + 0.07558366656303406, + 0.0036911526694893837, + 0.005526077002286911, + 0.004454519599676132, + 0.004550463519990444, + 0.007291737478226423, + 0.004160492680966854, + 0.004164418205618858, + 0.004110744688659906, + 0.01210648100823164, + 0.011624451726675034, + 0.0059144152328372, + 0.006807429250329733, + 0.013723128475248814, + 0.009680316783487797, + 0.0055161421187222, + 0.007035441696643829, + 0.013669755309820175, + 0.012454593554139137, + 0.011530180461704731, + 0.02554614096879959, + 0.012653155252337456, + 0.01499694399535656, + 0.01436647679656744, + 0.01037601102143526, + 0.015423179604113102, + 0.014038369059562683, + 0.010243387892842293, + 0.017045876011252403, + 0.012694956734776497, + 0.013463584706187248, + 0.011651000007987022, + 0.011928350664675236, + 0.018188072368502617, + 0.005675955675542355, + 0.004473140463232994, + 0.005307764280587435, + 0.0055578118190169334, + 0.022087063640356064, + 0.003636746434494853, + 0.007905948907136917, + 0.024381624534726143, + 0.010760348290205002, + 0.014111116528511047, + 0.010232912376523018, + 0.01262744702398777, + 0.010746598243713379, + 0.006121679674834013, + 0.022887064144015312, + 0.014464535750448704, + 0.0038891183212399483, + 0.013332733884453773, + 0.004410207737237215, + 0.004283627960830927, + 0.007391561754047871, + 0.005433422978967428, + 0.007508446462452412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.7578914165496826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "The camera executes a dramatic, low-angle, spiraling orbital shot, circling an upward flight of gleaming metallic robots as they blast upward in perfect unison from a circular, underground launch pad. Each robot is constructed of polished gunmetal-gray alloy to a mirror-like sheen, with a sleek, aerodynamic humanoid shape and integrated flight stabilizers on its back and limbs. Their featureless obsidian visors glow with an intense inner white light, and thin cobalt-blue energy conduits trace the seams of their armor plating. The action begins as they simultaneously ignite the powerful plasma thrusters beneath their feet. The initial blast of fiery blue energy shatters the concrete floor beneath them, sending a torrent of steam and debris cascading outward. As they ascend at breathtaking speed into a vast, dark silo, the camera stays in sync with them, perfectly executing a series of complex flips—a tight barrel roll before rapidly dispersing and regrouping into a precise diamond formation. The only sources of light are the robots’ own bright thruster plumes and glowing components, which cast harsh, rapidly changing highlights and deep, dancing shadows across their metallic bodies, creating an atmosphere of high contrast, intense dynamism, and breathtaking, overwhelming technological power.", + "seed": 12992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_77_seed_12992_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": -0.17689684107899667, + "vlm_logprob": -0.17689684107899667, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.17689684107899667, + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052216675132513046 + ], + "opposite_logprob_tokens": [ + -3.3022167682647705 + ], + "logprob_reward": -0.052216675132513046, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.052216675132513046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18212658166885376, + 0.13636453449726105, + 0.04674377664923668, + 0.0796799436211586, + 0.07625620812177658, + 0.05322130024433136, + 0.08334174007177353, + 0.05539260059595108, + 0.19045931100845337, + 0.0964139997959137 + ], + "credit_map": [ + 0.0294844638556242, + 0.016224943101406097, + 0.012676575221121311, + 0.019572677090764046, + 0.04097123444080353, + 0.016685819253325462, + 0.022896327078342438, + 0.023614542558789253, + 0.010704399086534977, + 0.017291652038693428, + 0.012337420135736465, + 0.014033781364560127, + 0.012811893597245216, + 0.033259909600019455, + 0.026569975540041924, + 0.009355505928397179, + 0.007145442999899387, + 0.006027580704540014, + 0.005763097200542688, + 0.00441509485244751, + 0.008236645720899105, + 0.007040165830403566, + 0.0043989066034555435, + 0.003716840874403715, + 0.010231801308691502, + 0.009819280356168747, + 0.008789854124188423, + 0.012273897416889668, + 0.014186397194862366, + 0.00821561086922884, + 0.007439952343702316, + 0.008723149076104164, + 0.009557467885315418, + 0.007268290966749191, + 0.005491574760526419, + 0.013685865327715874, + 0.011509381234645844, + 0.008484714664518833, + 0.010414179414510727, + 0.009844734333455563, + 0.004158818628638983, + 0.003229007823392749, + 0.0069589815102517605, + 0.004091487731784582, + 0.006053828168660402, + 0.01658260077238083, + 0.004371325485408306, + 0.007775248493999243, + 0.008950063958764076, + 0.010314780287444592, + 0.007030675653368235, + 0.012608685530722141, + 0.009051294066011906, + 0.013134634122252464, + 0.011877583339810371, + 0.010374021716415882, + 0.0064947581849992275, + 0.007258203346282244, + 0.007546849548816681, + 0.00617190869525075, + 0.007748859468847513, + 0.006632783450186253, + 0.007787457201629877, + 0.005751778371632099, + 0.02741292119026184, + 0.027140570804476738, + 0.030027536675333977, + 0.013998255133628845, + 0.03302760422229767, + 0.01679881475865841, + 0.020380398258566856, + 0.02167319692671299, + 0.013976674526929855, + 0.013301749713718891, + 0.012929385527968407, + 0.011277160607278347, + 0.01090301014482975, + 0.012278116308152676, + 0.009660657495260239, + 0.012087245471775532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.052216675132513046 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981521561741829 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.02981521561741829, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981521561741829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16172397136688232, + 0.5055388808250427, + 0.056393101811409, + 0.04192926362156868, + 0.03203168883919716, + 0.02964233048260212, + 0.04008645936846733, + 0.01938038505613804, + 0.05704522505402565, + 0.056228626519441605 + ], + "credit_map": [ + 0.04213187098503113, + 0.012336044572293758, + 0.008412330411374569, + 0.007082566153258085, + 0.050872448831796646, + 0.01919667422771454, + 0.01230947207659483, + 0.00938256923109293, + 0.037122346460819244, + 0.049775850027799606, + 0.03523184359073639, + 0.04195130243897438, + 0.048825912177562714, + 0.15289227664470673, + 0.1119578406214714, + 0.02778155915439129, + 0.007403148803859949, + 0.008432216942310333, + 0.004759014584124088, + 0.0082122553139925, + 0.010809092782437801, + 0.005657105706632137, + 0.006280928384512663, + 0.004839337896555662, + 0.00503547815605998, + 0.004536889493465424, + 0.005239746067672968, + 0.007688240613788366, + 0.006099754478782415, + 0.003211370436474681, + 0.004931384697556496, + 0.005186401307582855, + 0.00793962087482214, + 0.004250634461641312, + 0.002930421382188797, + 0.003307577921077609, + 0.0039549958892166615, + 0.0037307851016521454, + 0.002609123243018985, + 0.0033085301984101534, + 0.0029329771641641855, + 0.0019298242405056953, + 0.0025306991301476955, + 0.002121787518262863, + 0.004236527718603611, + 0.008435835130512714, + 0.0033463810104876757, + 0.004108297638595104, + 0.0038451545406132936, + 0.003430722514167428, + 0.005748978350311518, + 0.00707987928763032, + 0.004716727416962385, + 0.0042770323343575, + 0.004870219621807337, + 0.006117742974311113, + 0.002685871673747897, + 0.0021240708883851767, + 0.0029062561225146055, + 0.002005630638450384, + 0.0018177080200985074, + 0.0029952849727123976, + 0.0025551309809088707, + 0.0022904337383806705, + 0.009474861435592175, + 0.00972787942737341, + 0.00690965773537755, + 0.00434544961899519, + 0.01051999256014824, + 0.004531043116003275, + 0.0053662206046283245, + 0.0061701214872300625, + 0.007984943687915802, + 0.007704578805714846, + 0.008350702933967113, + 0.008533894084393978, + 0.007349108811467886, + 0.006137254182249308, + 0.004575246479362249, + 0.005592896603047848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02981521561741829 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636604942381382 + ], + "opposite_logprob_tokens": [ + -3.6513659954071045 + ], + "logprob_reward": -0.02636604942381382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636604942381382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14994429051876068, + 0.28162509202957153, + 0.06369809806346893, + 0.06840769201517105, + 0.058079794049263, + 0.06338327378034592, + 0.045501358807086945, + 0.03672840818762779, + 0.06297265738248825, + 0.1696593016386032 + ], + "credit_map": [ + 0.022387012839317322, + 0.014225292950868607, + 0.009971936233341694, + 0.01728787273168564, + 0.03559774532914162, + 0.014808370731770992, + 0.01692548207938671, + 0.018740588799118996, + 0.01809726655483246, + 0.027087971568107605, + 0.014824029058218002, + 0.0318734236061573, + 0.025666605681180954, + 0.07308415323495865, + 0.07284782826900482, + 0.01814381219446659, + 0.007813159376382828, + 0.009631261229515076, + 0.0065575698390603065, + 0.005144132766872644, + 0.009621113538742065, + 0.008856837637722492, + 0.009831896983087063, + 0.006242126226425171, + 0.009663756005465984, + 0.01388544775545597, + 0.006222120951861143, + 0.008593451231718063, + 0.012646082788705826, + 0.006250830367207527, + 0.006098408252000809, + 0.005047599319368601, + 0.013040650635957718, + 0.0045620836317539215, + 0.004843868315219879, + 0.008247945457696915, + 0.009289618581533432, + 0.005894096568226814, + 0.0055170622654259205, + 0.006684467662125826, + 0.005769992712885141, + 0.005182461813092232, + 0.0071607488207519054, + 0.00689640874043107, + 0.008029280230402946, + 0.012964733876287937, + 0.009889093227684498, + 0.0074905576184391975, + 0.00462149316444993, + 0.005288566928356886, + 0.004627482034265995, + 0.007556965108960867, + 0.005336718168109655, + 0.006149633321911097, + 0.006325327791273594, + 0.005595173221081495, + 0.005083908326923847, + 0.0042760055512189865, + 0.006315275561064482, + 0.004063985776156187, + 0.004621374420821667, + 0.005570529494434595, + 0.003720075823366642, + 0.0030772536993026733, + 0.008145863190293312, + 0.008572847582399845, + 0.010378367267549038, + 0.006191904190927744, + 0.009030934423208237, + 0.005663525313138962, + 0.00738895358517766, + 0.00760026415809989, + 0.021721964702010155, + 0.023006755858659744, + 0.03070412389934063, + 0.02412978559732437, + 0.018435195088386536, + 0.018416740000247955, + 0.01654086634516716, + 0.01670386828482151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.02636604942381382 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195146694779396 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195146694779396, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195146694779396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07102366536855698, + 0.14369435608386993, + 0.04119691252708435, + 0.05142662674188614, + 0.06332630664110184, + 0.1808420866727829, + 0.26855117082595825, + 0.07423878461122513, + 0.027566278353333473, + 0.07813382893800735 + ], + "credit_map": [ + 0.014104524627327919, + 0.005817062687128782, + 0.004109073895961046, + 0.0059929508715868, + 0.01690215989947319, + 0.011568780988454819, + 0.006176381837576628, + 0.006352728698402643, + 0.009334965609014034, + 0.01390780508518219, + 0.012754754163324833, + 0.017795979976654053, + 0.011175034567713737, + 0.03448474034667015, + 0.03447591885924339, + 0.00976516306400299, + 0.0063133020885288715, + 0.004743268247693777, + 0.0036393562331795692, + 0.003875268856063485, + 0.007089017424732447, + 0.006670908071100712, + 0.005178214516490698, + 0.0036875740624964237, + 0.006581975147128105, + 0.006300097797065973, + 0.00638691196218133, + 0.009092187508940697, + 0.006766706705093384, + 0.00433657830581069, + 0.0053855786100029945, + 0.006576588377356529, + 0.007212022785097361, + 0.00568052101880312, + 0.0062303137965500355, + 0.010745324194431305, + 0.009611373767256737, + 0.005780864041298628, + 0.00887142214924097, + 0.009194462560117245, + 0.015553475357592106, + 0.012029563076794147, + 0.016397828236222267, + 0.015831509605050087, + 0.022647855803370476, + 0.05607852712273598, + 0.015173343941569328, + 0.02712997794151306, + 0.03089987114071846, + 0.028519095852971077, + 0.02980129048228264, + 0.03702027350664139, + 0.03793567419052124, + 0.0338996946811676, + 0.03492637723684311, + 0.03554890677332878, + 0.009380139410495758, + 0.00878028105944395, + 0.01474108174443245, + 0.010496607981622219, + 0.0072207641787827015, + 0.011116011068224907, + 0.005825472995638847, + 0.006678431294858456, + 0.0038488958962261677, + 0.004053429234772921, + 0.005125562194734812, + 0.0028459185268729925, + 0.0037626740522682667, + 0.002205154625698924, + 0.002959120785817504, + 0.002765523735433817, + 0.01103539764881134, + 0.011948634870350361, + 0.02066311053931713, + 0.008040273562073708, + 0.00802441406995058, + 0.005282544530928135, + 0.007667049299925566, + 0.005472411401569843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.018195146694779396 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578911185264587 + ], + "opposite_logprob_tokens": [ + -0.6328911185264587 + ], + "logprob_reward": -0.7578911185264587, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578911185264587, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14093847572803497, + 0.15615040063858032, + 0.06227264180779457, + 0.10968329757452011, + 0.10422882437705994, + 0.07505141943693161, + 0.11342129111289978, + 0.10613947361707687, + 0.06761184334754944, + 0.06450232863426208 + ], + "credit_map": [ + 0.01953781582415104, + 0.015169053338468075, + 0.015034417621791363, + 0.017496507614850998, + 0.022238723933696747, + 0.018658284097909927, + 0.01473283488303423, + 0.01807083562016487, + 0.014041119255125523, + 0.017330460250377655, + 0.02252940461039543, + 0.01547933742403984, + 0.019387835636734962, + 0.03949658200144768, + 0.016949467360973358, + 0.010936182923614979, + 0.00811542384326458, + 0.007527095265686512, + 0.008250917308032513, + 0.005750581622123718, + 0.008950499817728996, + 0.00853579118847847, + 0.008517651818692684, + 0.006624680943787098, + 0.01359364204108715, + 0.010733832605183125, + 0.013740245252847672, + 0.02196960337460041, + 0.012883774936199188, + 0.008581624366343021, + 0.013784000650048256, + 0.014396576210856438, + 0.01982550136744976, + 0.008552853949368, + 0.008182251825928688, + 0.021910378709435463, + 0.01284386683255434, + 0.011900920420885086, + 0.010364504531025887, + 0.010648541152477264, + 0.007671790197491646, + 0.004284221678972244, + 0.009321206249296665, + 0.005212250165641308, + 0.007770848926156759, + 0.02414383739233017, + 0.00788391288369894, + 0.008763348683714867, + 0.009754371829330921, + 0.009287422522902489, + 0.009101612493395805, + 0.015669135376811028, + 0.0150392334908247, + 0.01640656217932701, + 0.019121987745165825, + 0.019040964543819427, + 0.012960391119122505, + 0.01282712072134018, + 0.01417213212698698, + 0.011022229678928852, + 0.012566180899739265, + 0.020879939198493958, + 0.010812929831445217, + 0.010898548178374767, + 0.008960994891822338, + 0.007354772184044123, + 0.013693331740796566, + 0.006740904878824949, + 0.006540616974234581, + 0.0065270280465483665, + 0.008519964292645454, + 0.00927423220127821, + 0.01107746921479702, + 0.011388459242880344, + 0.00856899842619896, + 0.006332388613373041, + 0.008288796059787273, + 0.00566951185464859, + 0.006647272501140833, + 0.006529434584081173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.7578911185264587 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "The camera executes a dramatic, low-angle, spiraling orbital shot, circling an upward flight of gleaming metallic robots as they blast upward in perfect unison from a circular, underground launch pad. Each robot is constructed of polished gunmetal-gray alloy to a mirror-like sheen, with a sleek, aerodynamic humanoid shape and integrated flight stabilizers on its back and limbs. Their featureless obsidian visors glow with an intense inner white light, and thin cobalt-blue energy conduits trace the seams of their armor plating. The action begins as they simultaneously ignite the powerful plasma thrusters beneath their feet. The initial blast of fiery blue energy shatters the concrete floor beneath them, sending a torrent of steam and debris cascading outward. As they ascend at breathtaking speed into a vast, dark silo, the camera stays in sync with them, perfectly executing a series of complex flips—a tight barrel roll before rapidly dispersing and regrouping into a precise diamond formation. The only sources of light are the robots’ own bright thruster plumes and glowing components, which cast harsh, rapidly changing highlights and deep, dancing shadows across their metallic bodies, creating an atmosphere of high contrast, intense dynamism, and breathtaking, overwhelming technological power.", + "seed": 12993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_77_seed_12993_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": -0.10966681716963649, + "vlm_logprob": -0.10966681716963649, + "vlm_yes_no_margin": 3.224999988079071, + "vlm_token_credit": -0.10966681716963649, + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04571854695677757 + ], + "opposite_logprob_tokens": [ + -3.2957184314727783 + ], + "logprob_reward": -0.04571854695677757, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04571854695677757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1477748602628708, + 0.07283473014831543, + 0.06944791227579117, + 0.05128372088074684, + 0.15772289037704468, + 0.22520406544208527, + 0.04760929197072983, + 0.07230658084154129, + 0.04660869017243385, + 0.10920725017786026 + ], + "credit_map": [ + 0.013984204269945621, + 0.014502021484076977, + 0.02031416818499565, + 0.019256798550486565, + 0.014587200246751308, + 0.01768535189330578, + 0.01974981464445591, + 0.02769530564546585, + 0.009379069320857525, + 0.005710006691515446, + 0.010535292327404022, + 0.013754000887274742, + 0.008603758178651333, + 0.010869476944208145, + 0.008271418511867523, + 0.005711712874472141, + 0.007759573869407177, + 0.007761664222925901, + 0.008399153128266335, + 0.011966205202043056, + 0.013551538810133934, + 0.008615576662123203, + 0.0061147320084273815, + 0.005279474891722202, + 0.006136998534202576, + 0.00819791667163372, + 0.004222225863486528, + 0.006560689769685268, + 0.00669383630156517, + 0.006309150252491236, + 0.008002403192222118, + 0.005160500295460224, + 0.01917545683681965, + 0.015292972326278687, + 0.016862640157341957, + 0.021165603771805763, + 0.017540927976369858, + 0.02257634326815605, + 0.028187468647956848, + 0.016921473667025566, + 0.03167857602238655, + 0.020653050392866135, + 0.03738074377179146, + 0.02720765583217144, + 0.025340627878904343, + 0.023222215473651886, + 0.03436458855867386, + 0.025356611236929893, + 0.004056548699736595, + 0.004471583757549524, + 0.0047932215966284275, + 0.003941723611205816, + 0.0039484393782913685, + 0.007263776380568743, + 0.005261938087642193, + 0.013872065581381321, + 0.009090686216950417, + 0.007876920513808727, + 0.008386751636862755, + 0.008964941836893559, + 0.0072525194846093655, + 0.01174861378967762, + 0.008295390754938126, + 0.01069075521081686, + 0.0044042919762432575, + 0.003987782634794712, + 0.007674132473766804, + 0.004219661001116037, + 0.006998294964432716, + 0.006533374078571796, + 0.0054396954365074635, + 0.00735145527869463, + 0.01330543216317892, + 0.014547575265169144, + 0.013819025829434395, + 0.013464665971696377, + 0.014021665789186954, + 0.011655815877020359, + 0.013125913217663765, + 0.015267151407897472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.04571854695677757 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054961517453193665 + ], + "opposite_logprob_tokens": [ + -2.9299614429473877 + ], + "logprob_reward": -0.054961517453193665, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054961517453193665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1247531846165657, + 0.041568268090486526, + 0.08413249254226685, + 0.09609190374612808, + 0.06409060955047607, + 0.14211691915988922, + 0.18684574961662292, + 0.08597684651613235, + 0.0802282840013504, + 0.09419573098421097 + ], + "credit_map": [ + 0.015585297718644142, + 0.014494235627353191, + 0.012561789713799953, + 0.013874956406652927, + 0.01660054735839367, + 0.013143932446837425, + 0.018581567332148552, + 0.01991085894405842, + 0.004911694210022688, + 0.0033686126116663218, + 0.011571889743208885, + 0.004979781340807676, + 0.003871660213917494, + 0.007094415370374918, + 0.0030803410336375237, + 0.0026898717042058706, + 0.009458702057600021, + 0.007677385117858648, + 0.010628067888319492, + 0.010324137285351753, + 0.01602742075920105, + 0.011992966756224632, + 0.009940596297383308, + 0.008083216845989227, + 0.007745690643787384, + 0.01006264053285122, + 0.008942329324781895, + 0.012790417298674583, + 0.017180534079670906, + 0.01640351675450802, + 0.012041997164487839, + 0.010924777016043663, + 0.0078008146956563, + 0.005981603171676397, + 0.00782644934952259, + 0.00710415281355381, + 0.006242819130420685, + 0.00888735055923462, + 0.010777517221868038, + 0.0094698965549469, + 0.021906904876232147, + 0.01127608772367239, + 0.022608524188399315, + 0.015507050789892673, + 0.020878871902823448, + 0.016000986099243164, + 0.018538812175393105, + 0.015399686060845852, + 0.0182565376162529, + 0.012950843200087547, + 0.012640670873224735, + 0.022036781534552574, + 0.024379659444093704, + 0.026970822364091873, + 0.020170411095023155, + 0.04944002255797386, + 0.01235970389097929, + 0.007831995375454426, + 0.010335786268115044, + 0.008441798388957977, + 0.010278563015162945, + 0.013594443909823895, + 0.010598666034638882, + 0.012535889632999897, + 0.009924771264195442, + 0.005612572189420462, + 0.016991671174764633, + 0.01002347469329834, + 0.009355328977108002, + 0.008620412088930607, + 0.008612556383013725, + 0.011087498627603054, + 0.012673418037593365, + 0.012852328829467297, + 0.012033089995384216, + 0.011970766820013523, + 0.011299225501716137, + 0.010749656707048416, + 0.01142624020576477, + 0.011190999299287796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.054961517453193665 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.18685247004032135, + 0.2941941022872925, + 0.07794467359781265, + 0.035308580845594406, + 0.03996659070253372, + 0.08867256343364716, + 0.06267179548740387, + 0.0669439509510994, + 0.10224483907222748, + 0.04520035907626152 + ], + "credit_map": [ + 0.013700014911592007, + 0.023474112153053284, + 0.018716810271143913, + 0.0263872928917408, + 0.02084171213209629, + 0.019030028954148293, + 0.027249040082097054, + 0.03745346516370773, + 0.024826517328619957, + 0.014709494076669216, + 0.10954231023788452, + 0.016405832022428513, + 0.01788076013326645, + 0.08671695739030838, + 0.014573940075933933, + 0.009538319893181324, + 0.007834107615053654, + 0.010707169771194458, + 0.006644826382398605, + 0.009984725154936314, + 0.012855827808380127, + 0.008655759505927563, + 0.011666025966405869, + 0.009596237912774086, + 0.00288196187466383, + 0.004050559364259243, + 0.0029698885045945644, + 0.0040443819016218185, + 0.004798654466867447, + 0.0060271453112363815, + 0.005599565804004669, + 0.004936421290040016, + 0.00690446188673377, + 0.003734721802175045, + 0.0036223279312253, + 0.004274964798241854, + 0.0045572733506560326, + 0.006319273728877306, + 0.00594890583306551, + 0.004604661837220192, + 0.010829810984432697, + 0.009730737656354904, + 0.010204467922449112, + 0.010535060428082943, + 0.013723046518862247, + 0.010837850160896778, + 0.013763432390987873, + 0.009048156440258026, + 0.005388213787227869, + 0.006023301277309656, + 0.006555055268108845, + 0.008624522015452385, + 0.010949886403977871, + 0.00790198240429163, + 0.007513647899031639, + 0.009715186432003975, + 0.006417213939130306, + 0.007304142229259014, + 0.007577199023216963, + 0.010376590304076672, + 0.008989598602056503, + 0.008458955213427544, + 0.009566365741193295, + 0.008253886364400387, + 0.009974159300327301, + 0.01007203757762909, + 0.02024662122130394, + 0.008697994984686375, + 0.018289711326360703, + 0.014947125688195229, + 0.00932713970541954, + 0.010690048336982727, + 0.005121802445501089, + 0.007520715706050396, + 0.004681587219238281, + 0.006537905894219875, + 0.005651515442878008, + 0.005141787696629763, + 0.005734322592616081, + 0.004810723941773176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763084227219224 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.00763084227219224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763084227219224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15683263540267944, + 0.11032216995954514, + 0.053612712770700455, + 0.06806573271751404, + 0.030555685982108116, + 0.059974320232868195, + 0.04404034465551376, + 0.14313356578350067, + 0.1946842074394226, + 0.13877859711647034 + ], + "credit_map": [ + 0.012704234570264816, + 0.01955769769847393, + 0.013817498460412025, + 0.027134282514452934, + 0.01921766623854637, + 0.013494531624019146, + 0.02191426232457161, + 0.02899244986474514, + 0.00958233792334795, + 0.007838810794055462, + 0.03318115696310997, + 0.009222335182130337, + 0.008877985179424286, + 0.02497393637895584, + 0.008875755593180656, + 0.007769863121211529, + 0.0066139050759375095, + 0.00688913743942976, + 0.006391164381057024, + 0.0067339395172894, + 0.009483825415372849, + 0.0070663681253790855, + 0.00575909623876214, + 0.004675277043133974, + 0.008703089319169521, + 0.010454295203089714, + 0.005329744424670935, + 0.005911647342145443, + 0.010686243884265423, + 0.010308511555194855, + 0.00873924046754837, + 0.007932965643703938, + 0.0050130728632211685, + 0.0028200882952660322, + 0.003119108034297824, + 0.003403135808184743, + 0.004201867152005434, + 0.004201575648039579, + 0.0052287522703409195, + 0.0025680873077362776, + 0.008997282013297081, + 0.004443122074007988, + 0.010007010772824287, + 0.00833972916007042, + 0.009121323935687542, + 0.004614435601979494, + 0.00855876412242651, + 0.005892653949558735, + 0.004101223777979612, + 0.004255272448062897, + 0.004339126870036125, + 0.006678301375359297, + 0.006589186377823353, + 0.005232531111687422, + 0.004673530347645283, + 0.008171170949935913, + 0.01656998135149479, + 0.022217830643057823, + 0.014461004175245762, + 0.016806695610284805, + 0.014128195121884346, + 0.023021506145596504, + 0.016349632292985916, + 0.01957871951162815, + 0.01741008087992668, + 0.020344287157058716, + 0.038394227623939514, + 0.017125243321061134, + 0.03394203633069992, + 0.02797425538301468, + 0.019140634685754776, + 0.02035343088209629, + 0.021676385775208473, + 0.011855467222630978, + 0.02423979341983795, + 0.017935488373041153, + 0.013102304190397263, + 0.014446874149143696, + 0.012353644706308842, + 0.023168636485934258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.00763084227219224 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893558740615845 + ], + "opposite_logprob_tokens": [ + -1.0539355278015137 + ], + "logprob_reward": -0.42893558740615845, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42893558740615845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12944026291370392, + 0.036595847457647324, + 0.07178439199924469, + 0.061321478337049484, + 0.061815738677978516, + 0.1219913512468338, + 0.08346235752105713, + 0.1980644166469574, + 0.14289450645446777, + 0.09262959659099579 + ], + "credit_map": [ + 0.016135670244693756, + 0.015209187753498554, + 0.014771302230656147, + 0.013899735175073147, + 0.01277962327003479, + 0.013909597881138325, + 0.016069471836090088, + 0.026665683835744858, + 0.0039271404966712, + 0.003927342593669891, + 0.009632131084799767, + 0.0029947219882160425, + 0.0037887084763497114, + 0.0066870530135929585, + 0.0028604252729564905, + 0.0027783261612057686, + 0.00746729364618659, + 0.009875763207674026, + 0.007608124986290932, + 0.009669041261076927, + 0.013460400514304638, + 0.008171960711479187, + 0.009172433987259865, + 0.006359371356666088, + 0.007208117749541998, + 0.010681487619876862, + 0.005856804084032774, + 0.005958249792456627, + 0.008700262755155563, + 0.007945037446916103, + 0.00861549936234951, + 0.006356014870107174, + 0.007683406118303537, + 0.007049450185149908, + 0.00988294929265976, + 0.008165538311004639, + 0.006871287245303392, + 0.008299085311591625, + 0.00793696939945221, + 0.005927048157900572, + 0.018052218481898308, + 0.011660792864859104, + 0.018977832049131393, + 0.013068773783743382, + 0.01914171688258648, + 0.01318211480975151, + 0.016022877767682076, + 0.011885021813213825, + 0.007162717171013355, + 0.008508237078785896, + 0.008704022504389286, + 0.0127746956422925, + 0.01502998173236847, + 0.010526490397751331, + 0.009121435694396496, + 0.01163477823138237, + 0.022379182279109955, + 0.01908816583454609, + 0.02124135196208954, + 0.03025791607797146, + 0.029671231284737587, + 0.024027328938245773, + 0.02546527050435543, + 0.025933973491191864, + 0.010045320726931095, + 0.01647930033504963, + 0.02519533783197403, + 0.012937912717461586, + 0.024881809949874878, + 0.02050914242863655, + 0.01592954620718956, + 0.01691613532602787, + 0.008807668462395668, + 0.013804699294269085, + 0.008984152227640152, + 0.01458209939301014, + 0.009734278544783592, + 0.009661031886935234, + 0.014101160690188408, + 0.01295450422912836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.42893558740615845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera executes a dramatic, low-angle, spiraling orbital shot, circling an upward flight of gleaming metallic robots as they blast upward in perfect unison from a circular, underground launch pad. Each robot is constructed of polished gunmetal-gray alloy to a mirror-like sheen, with a sleek, aerodynamic humanoid shape and integrated flight stabilizers on its back and limbs. Their featureless obsidian visors glow with an intense inner white light, and thin cobalt-blue energy conduits trace the seams of their armor plating. The action begins as they simultaneously ignite the powerful plasma thrusters beneath their feet. The initial blast of fiery blue energy shatters the concrete floor beneath them, sending a torrent of steam and debris cascading outward. As they ascend at breathtaking speed into a vast, dark silo, the camera stays in sync with them, perfectly executing a series of complex flips—a tight barrel roll before rapidly dispersing and regrouping into a precise diamond formation. The only sources of light are the robots’ own bright thruster plumes and glowing components, which cast harsh, rapidly changing highlights and deep, dancing shadows across their metallic bodies, creating an atmosphere of high contrast, intense dynamism, and breathtaking, overwhelming technological power.", + "seed": 12994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_14/index_77_seed_12994_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": -0.21136325523257254, + "vlm_logprob": -0.21136325523257254, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.21136325523257254, + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20206105709075928 + ], + "opposite_logprob_tokens": [ + -1.8270610570907593 + ], + "logprob_reward": -0.20206105709075928, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20206105709075928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04178435727953911, + 0.024884559214115143, + 0.027266614139080048, + 0.687770664691925, + 0.009748689830303192, + 0.06446568667888641, + 0.026403965428471565, + 0.051556844264268875, + 0.031116198748350143, + 0.03500239923596382 + ], + "credit_map": [ + 0.006367096211761236, + 0.006884316448122263, + 0.0038654482923448086, + 0.004881358239799738, + 0.005032780114561319, + 0.004703463986515999, + 0.004956149961799383, + 0.005093750543892384, + 0.003451680066064, + 0.005610843654721975, + 0.0016763334861025214, + 0.0020702588371932507, + 0.0035716453567147255, + 0.00346190994605422, + 0.0017514475621283054, + 0.003290443681180477, + 0.0037648871075361967, + 0.004573410842567682, + 0.004284115973860025, + 0.0024077831767499447, + 0.0030844570137560368, + 0.003631473518908024, + 0.002691476373001933, + 0.0028290124610066414, + 0.05528253689408302, + 0.021094439551234245, + 0.023550203070044518, + 0.027893254533410072, + 0.059173330664634705, + 0.23612746596336365, + 0.047836244106292725, + 0.21681326627731323, + 0.0008359689963981509, + 0.0007504193927161396, + 0.0010274136438965797, + 0.0015626682434231043, + 0.0016395078273490071, + 0.0007913470035418868, + 0.0015486025949940085, + 0.001592762884683907, + 0.013014824129641056, + 0.0029564059805125, + 0.005669944453984499, + 0.008208474144339561, + 0.010584786534309387, + 0.0038295076228678226, + 0.0077749802730977535, + 0.01242676842957735, + 0.0032627396285533905, + 0.0013223689747974277, + 0.0015388975152745843, + 0.008214563131332397, + 0.0030498001724481583, + 0.0015290118753910065, + 0.003911179956048727, + 0.003575408598408103, + 0.006975335534662008, + 0.0030816129874438047, + 0.003664277493953705, + 0.008017380721867085, + 0.010846743360161781, + 0.002998120617121458, + 0.009811267256736755, + 0.006162112113088369, + 0.003493820782750845, + 0.0027461363933980465, + 0.0032897107303142548, + 0.003208567388355732, + 0.004267039243131876, + 0.004165515769273043, + 0.004865078255534172, + 0.005080333445221186, + 0.004378046374768019, + 0.002614010591059923, + 0.00718273688107729, + 0.004179114475846291, + 0.003164363093674183, + 0.00443099532276392, + 0.0038084429688751698, + 0.005244692321866751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.20206105709075928 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812770918011665 + ], + "opposite_logprob_tokens": [ + -3.2881276607513428 + ], + "logprob_reward": -0.03812770918011665, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812770918011665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11277052015066147, + 0.07143601775169373, + 0.25237417221069336, + 0.21890577673912048, + 0.012662874534726143, + 0.10017342120409012, + 0.018637007102370262, + 0.06191995367407799, + 0.0709294006228447, + 0.08019078522920609 + ], + "credit_map": [ + 0.012482254765927792, + 0.029392637312412262, + 0.006461948622018099, + 0.009361489675939083, + 0.01618306338787079, + 0.014150415547192097, + 0.012007719837129116, + 0.012730984017252922, + 0.010140915401279926, + 0.017252760007977486, + 0.005702804308384657, + 0.008039485663175583, + 0.00911048986017704, + 0.009158548898994923, + 0.005646980833262205, + 0.006384034175425768, + 0.03016204573214054, + 0.04288335517048836, + 0.03199027106165886, + 0.026913048699498177, + 0.03545043244957924, + 0.038866255432367325, + 0.02799108810722828, + 0.018117673695087433, + 0.01785178855061531, + 0.006037845741957426, + 0.009426183998584747, + 0.009549692273139954, + 0.016533227637410164, + 0.07554357498884201, + 0.015142034739255905, + 0.06882143020629883, + 0.0012456827098503709, + 0.0008867086726240814, + 0.0018149983370676637, + 0.003073212457820773, + 0.0024067757185548544, + 0.0008571898797526956, + 0.0012153210118412971, + 0.0011629854561761022, + 0.017503980547189713, + 0.004958584904670715, + 0.011426901444792747, + 0.015079788863658905, + 0.014668961055576801, + 0.011096031405031681, + 0.015026690438389778, + 0.010412482544779778, + 0.0026486734859645367, + 0.0009555231663398445, + 0.0013112021842971444, + 0.0036345093976706266, + 0.003976975101977587, + 0.0012537834700196981, + 0.0028068104293197393, + 0.002049530390650034, + 0.0073919715359807014, + 0.0036222899798303843, + 0.01005072332918644, + 0.01069447211921215, + 0.009384089149534702, + 0.004171045497059822, + 0.00922953337430954, + 0.00737582566216588, + 0.006217652466148138, + 0.0066823819652199745, + 0.010756241157650948, + 0.010331491008400917, + 0.009844313375651836, + 0.007161598652601242, + 0.007660625036805868, + 0.012275096960365772, + 0.009854084812104702, + 0.005379950627684593, + 0.01060433778911829, + 0.018446635454893112, + 0.007505280897021294, + 0.009986760094761848, + 0.008454523049294949, + 0.009959210641682148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.03812770918011665 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607881247997284 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607881247997284, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607881247997284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0374743826687336, + 0.046081461012363434, + 0.07283568382263184, + 0.7117100358009338, + 0.01526794396340847, + 0.04608634114265442, + 0.011146309785544872, + 0.024653911590576172, + 0.022296661511063576, + 0.01244722492992878 + ], + "credit_map": [ + 0.0043673268519341946, + 0.008484162390232086, + 0.002555222250521183, + 0.0032142712734639645, + 0.005332842003554106, + 0.0039017517119646072, + 0.003991963807493448, + 0.005626838654279709, + 0.004986835643649101, + 0.008310096338391304, + 0.004938665311783552, + 0.0059461575001478195, + 0.005152879282832146, + 0.0054315607994794846, + 0.003445559414103627, + 0.007869703695178032, + 0.009072832763195038, + 0.00792177114635706, + 0.012137453071773052, + 0.007314580958336592, + 0.0070580109022557735, + 0.008869391866028309, + 0.012527788989245892, + 0.007933849468827248, + 0.057578686624765396, + 0.019593263044953346, + 0.027971824631094933, + 0.029851093888282776, + 0.05700082331895828, + 0.24497167766094208, + 0.04963476210832596, + 0.2251078188419342, + 0.0021450649946928024, + 0.001354617765173316, + 0.001243576523847878, + 0.0023653972893953323, + 0.003392439102753997, + 0.0009812507778406143, + 0.0017608082853257656, + 0.0020247879438102245, + 0.0070008449256420135, + 0.0019691535271704197, + 0.007499557454138994, + 0.008701303042471409, + 0.005894877016544342, + 0.005546944215893745, + 0.005099928472191095, + 0.004373726435005665, + 0.0019820011220872402, + 0.0006697524804621935, + 0.0009323296835646033, + 0.0018461729632690549, + 0.0012641469947993755, + 0.0011220393935218453, + 0.001243134611286223, + 0.002086732303723693, + 0.003252460155636072, + 0.0013876761076971889, + 0.001710623037070036, + 0.0043082041665911674, + 0.00483732670545578, + 0.0013635744107887149, + 0.005028209183365107, + 0.002765836426988244, + 0.002351237926632166, + 0.0019947481341660023, + 0.0032596280798316, + 0.002034934004768729, + 0.0032015678007155657, + 0.0026846083346754313, + 0.00272609805688262, + 0.004043837543576956, + 0.0012547650840133429, + 0.0008366168476641178, + 0.0020515969954431057, + 0.0022498934995383024, + 0.0010709691559895873, + 0.001863367622718215, + 0.001607681275345385, + 0.0015123329358175397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.020607881247997284 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810866177082062 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.03810866177082062, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810866177082062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05341818556189537, + 0.038975004106760025, + 0.047084491699934006, + 0.6198348999023438, + 0.018002206459641457, + 0.07827270776033401, + 0.03210338577628136, + 0.03187679871916771, + 0.05275772139430046, + 0.027674468234181404 + ], + "credit_map": [ + 0.006859161891043186, + 0.009942720644176006, + 0.0033869845792651176, + 0.004990709945559502, + 0.010115573182702065, + 0.00600128248333931, + 0.004766774829477072, + 0.007354981731623411, + 0.004631440620869398, + 0.006323428358882666, + 0.004714504349976778, + 0.005873699672520161, + 0.004165349993854761, + 0.003865259699523449, + 0.0027483401354402304, + 0.0066529824398458, + 0.005678333342075348, + 0.00724439974874258, + 0.006486544851213694, + 0.004743597004562616, + 0.005942581221461296, + 0.006391999311745167, + 0.006116320379078388, + 0.004480716306716204, + 0.05017707869410515, + 0.016609370708465576, + 0.022850463166832924, + 0.02605920471251011, + 0.04945388808846474, + 0.21325762569904327, + 0.04380885139107704, + 0.19761843979358673, + 0.0019667574670165777, + 0.0010953806340694427, + 0.0019370103254914284, + 0.0035352406557649374, + 0.004913054872304201, + 0.0013582768151536584, + 0.0017130870837718248, + 0.0014833981404080987, + 0.011926564387977123, + 0.00370370433665812, + 0.007891791872680187, + 0.011200509034097195, + 0.013920115306973457, + 0.00911424495279789, + 0.012035797350108624, + 0.008479982614517212, + 0.0042838118970394135, + 0.001777553465217352, + 0.0017285492504015565, + 0.008999309502542019, + 0.004744295496493578, + 0.0022290002088993788, + 0.00472779618576169, + 0.0036130687221884727, + 0.003997753374278545, + 0.0020963104907423258, + 0.00541030615568161, + 0.006259097717702389, + 0.004290416371077299, + 0.0018631487619131804, + 0.0044577172957360744, + 0.00350204948335886, + 0.004753270652145147, + 0.00497741112485528, + 0.008189015090465546, + 0.006683190818876028, + 0.005809719208627939, + 0.0063720326870679855, + 0.007998619228601456, + 0.007974463514983654, + 0.0031436041463166475, + 0.002061105566099286, + 0.004520909860730171, + 0.004585764836519957, + 0.0029525193385779858, + 0.0042982627637684345, + 0.0026082247495651245, + 0.003504078835248947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.03810866177082062 + }, + { + "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\": \"Are the main subjects gleaming metallic robots with a sleek, aerodynamic humanoid shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a circular underground launch pad inside a vast, dark silo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the robots blast upward in perfect unison, igniting powerful plasma thrusters that shatter the concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dramatic, low-angle, spiraling orbital shot circling the robots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the robots form a precise diamond formation after the camera performs a tight barrel roll and disperses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.757910966873169 + ], + "opposite_logprob_tokens": [ + -0.632910966873169 + ], + "logprob_reward": -0.757910966873169, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.757910966873169, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.058766476809978485, + 0.05084942653775215, + 0.038135603070259094, + 0.527127206325531, + 0.04052026569843292, + 0.12541572749614716, + 0.032712336629629135, + 0.04065687209367752, + 0.04702301695942879, + 0.03879309445619583 + ], + "credit_map": [ + 0.0071205440908670425, + 0.011564239859580994, + 0.004479979630559683, + 0.00663005793467164, + 0.00799256470054388, + 0.0056051090359687805, + 0.006588528398424387, + 0.00878545269370079, + 0.006688663735985756, + 0.01172386109828949, + 0.005620721261948347, + 0.005891639739274979, + 0.0065736849792301655, + 0.0059033893048763275, + 0.0033759684301912785, + 0.0050714993849396706, + 0.004614833742380142, + 0.006126771215349436, + 0.00625778129324317, + 0.0033236502204090357, + 0.004351985175162554, + 0.004918611608445644, + 0.004424090031534433, + 0.004117881413549185, + 0.043104030191898346, + 0.013657419010996819, + 0.01941622793674469, + 0.021317528560757637, + 0.04143182188272476, + 0.18342585861682892, + 0.036850251257419586, + 0.16792407631874084, + 0.004780760500580072, + 0.0021105087362229824, + 0.0030502972658723593, + 0.009532267227768898, + 0.013605057261884212, + 0.0021703557576984167, + 0.002570147393271327, + 0.002700874349102378, + 0.01732330024242401, + 0.0050198412500321865, + 0.01621721126139164, + 0.015043118968605995, + 0.02160867117345333, + 0.01095118559896946, + 0.019353071227669716, + 0.01989932358264923, + 0.004146281164139509, + 0.002180412644520402, + 0.0026239044964313507, + 0.009893883019685745, + 0.004009267780929804, + 0.0017629930516704917, + 0.004808944184333086, + 0.0032866483088582754, + 0.005430274643003941, + 0.0023643954191356897, + 0.003990217577666044, + 0.006099481135606766, + 0.006811439525336027, + 0.002266343217343092, + 0.008593997918069363, + 0.005100722424685955, + 0.0037644815165549517, + 0.004105019383132458, + 0.0065643275156617165, + 0.006007966585457325, + 0.007097999099642038, + 0.004706148523837328, + 0.009453676640987396, + 0.005323395598679781, + 0.005163109861314297, + 0.0033565203193575144, + 0.007464323192834854, + 0.005333029665052891, + 0.0029510087333619595, + 0.0057967547327280045, + 0.003963340539485216, + 0.004765008110553026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.757910966873169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_tensor([2137, 2137])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_tensor([2137, 2137])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..54b3e2d3e6e7b914a601f45de219538dc127342a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_tensor([2137, 2137])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up depicts a mesmerizing vertical journey as a tiny bubble ascends through a crystal-clear, flaky mass of carbonated liquid contained in a sleek, minimalist glass. The scene is lit from the side by a soft, diffuse light source, casting clean, bright highlights along the curve of the glass and creating a luminous, almost ethereal quality. As the bubble begins its ascent from the bottom of the frame, the camera carefully tracks it; initially a perfect, shimmering microsphere, it subtly jiggles as it rises, its shape delicately deforming and vibrating under the pressure of the fluid. Its transparent membrane captures the light, refracting it into miniature rainbows with shimmering edges, while the clear liquid it passes through amplifies the softly blurred neutral background into a gentle bokeh. The bubble's steady, unwavering upward motion creates a tranquil and hypnotic atmosphere, focusing the viewer's full attention on this simple, beautiful phenomenon of physics and light as it moves from the bottom to the top of the frame.", + "seed": 4452, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_2137_seed_4452_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": -0.018123411992564797, + "vlm_logprob": -0.018123411992564797, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.018123411992564797, + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.01015295647084713 + ], + "opposite_logprob_tokens": [ + -6.510152816772461 + ], + "logprob_reward": -0.01015295647084713, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.01015295647084713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.124728262424469, + 0.035104900598526, + 0.12721385061740875, + 0.028288275003433228, + 0.15013696253299713, + 0.03830600157380104, + 0.01787756197154522, + 0.02468324638903141, + 0.29929062724113464, + 0.1543702781200409 + ], + "credit_map": [ + 0.021111810579895973, + 0.021873412653803825, + 0.012598976492881775, + 0.011652485467493534, + 0.012900058180093765, + 0.016089515760540962, + 0.014401200227439404, + 0.014100809581577778, + 0.004411371890455484, + 0.005248837172985077, + 0.0034557159524410963, + 0.004966116510331631, + 0.003961401991546154, + 0.004427255596965551, + 0.005456368904560804, + 0.0031778367701917887, + 0.010424547828733921, + 0.016797170042991638, + 0.02619912289083004, + 0.017093917354941368, + 0.012735971249639988, + 0.01117737963795662, + 0.015738127753138542, + 0.017047615721821785, + 0.004306577146053314, + 0.0030535231344401836, + 0.004611914046108723, + 0.003309175604954362, + 0.0024512489326298237, + 0.0037634652107954025, + 0.0028056420851498842, + 0.003986729774624109, + 0.009330682456493378, + 0.008566964417696, + 0.015538771636784077, + 0.008847844786942005, + 0.006488322746008635, + 0.05764980986714363, + 0.0058900341391563416, + 0.03782454878091812, + 0.0029629457276314497, + 0.0045563154853880405, + 0.00357522489503026, + 0.003518468700349331, + 0.005060098133981228, + 0.0069116284139454365, + 0.0036181353498250246, + 0.008103186264634132, + 0.002255385974422097, + 0.0027823352720588446, + 0.002798805944621563, + 0.0023534202482551336, + 0.001708249095827341, + 0.002202413510531187, + 0.0018179492326453328, + 0.0019590030424296856, + 0.0019196558278054, + 0.004053094424307346, + 0.0026128552854061127, + 0.002607160247862339, + 0.004451891873031855, + 0.0033189018722623587, + 0.002041393890976906, + 0.0036782920360565186, + 0.02380618080496788, + 0.06877879053354263, + 0.028629116714000702, + 0.02869829349219799, + 0.0329773873090744, + 0.03703358769416809, + 0.023729003965854645, + 0.05563829094171524, + 0.017733432352542877, + 0.018078839406371117, + 0.024840455502271652, + 0.01538597047328949, + 0.023039668798446655, + 0.018989771604537964, + 0.015402715653181076, + 0.020899426192045212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.01015295647084713 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.012556072324514389 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556072324514389, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556072324514389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19151680171489716, + 0.08927842974662781, + 0.14283061027526855, + 0.018220126628875732, + 0.18020084500312805, + 0.114427849650383, + 0.025559071451425552, + 0.01447973120957613, + 0.09938782453536987, + 0.12409868836402893 + ], + "credit_map": [ + 0.02291232906281948, + 0.040339820086956024, + 0.016923580318689346, + 0.013509087264537811, + 0.03143869340419769, + 0.02373656816780567, + 0.019518105313181877, + 0.023138616234064102, + 0.007674399763345718, + 0.014214660972356796, + 0.00863774586468935, + 0.0168340764939785, + 0.009668363258242607, + 0.009346597827970982, + 0.016042569652199745, + 0.006860018242150545, + 0.01293977815657854, + 0.018275829032063484, + 0.028960373252630234, + 0.01709950901567936, + 0.014924149960279465, + 0.013827764429152012, + 0.015911031514406204, + 0.020892178639769554, + 0.002639801474288106, + 0.0017238243017345667, + 0.0028199332300573587, + 0.0021722072269767523, + 0.002110666362568736, + 0.0023403188679367304, + 0.0017809367273002863, + 0.0026324389036744833, + 0.01030708011239767, + 0.008069372735917568, + 0.0183021891862154, + 0.009351812303066254, + 0.006242330186069012, + 0.08219430595636368, + 0.004724394995719194, + 0.04100935533642769, + 0.01175124291330576, + 0.008108355104923248, + 0.008605767041444778, + 0.006318701896816492, + 0.005695483181625605, + 0.036033995449543, + 0.0066055189818143845, + 0.03130877763032913, + 0.00350900087505579, + 0.0027804551646113396, + 0.004432988818734884, + 0.0029153688810765743, + 0.002887527458369732, + 0.004132206551730633, + 0.0021238410845398903, + 0.0027776837814599276, + 0.0014194310642778873, + 0.0015357235679402947, + 0.0027331202290952206, + 0.0014374584425240755, + 0.0011671723332256079, + 0.0030921807046979666, + 0.0015516867861151695, + 0.001542958663776517, + 0.007046030834317207, + 0.01690364070236683, + 0.010840787552297115, + 0.008723821491003036, + 0.009874674491584301, + 0.01077007595449686, + 0.01119360513985157, + 0.024035191163420677, + 0.012348213233053684, + 0.014720499515533447, + 0.013747543096542358, + 0.01419377326965332, + 0.023331794887781143, + 0.01505611464381218, + 0.013513334095478058, + 0.017187414690852165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.012556072324514389 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.003197202691808343 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.003197202691808343, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197202691808343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11293602734804153, + 0.11589362472295761, + 0.06336900591850281, + 0.039378389716148376, + 0.24095121026039124, + 0.05658316612243652, + 0.06788795441389084, + 0.0637400820851326, + 0.14643926918506622, + 0.09282130002975464 + ], + "credit_map": [ + 0.013467665761709213, + 0.027135074138641357, + 0.008682513609528542, + 0.0090127969160676, + 0.019257722422480583, + 0.012941177003085613, + 0.01093731913715601, + 0.011501763015985489, + 0.011897617019712925, + 0.0166949350386858, + 0.010961038991808891, + 0.020511656999588013, + 0.011773170903325081, + 0.014838176779448986, + 0.020425228402018547, + 0.008791802451014519, + 0.004999092314392328, + 0.009917271323502064, + 0.008558783680200577, + 0.008224084042012691, + 0.007886909879744053, + 0.006614162120968103, + 0.009349143132567406, + 0.007819565013051033, + 0.003952132072299719, + 0.00398994330316782, + 0.0076406714506447315, + 0.004471178166568279, + 0.0037494581192731857, + 0.004811447113752365, + 0.0035028692800551653, + 0.0072606890462338924, + 0.012891153804957867, + 0.009749723598361015, + 0.028420133516192436, + 0.01246548444032669, + 0.007116336841136217, + 0.10367550700902939, + 0.005958225578069687, + 0.06067465990781784, + 0.006518268026411533, + 0.006999907083809376, + 0.0063910167664289474, + 0.005906160455197096, + 0.009801196865737438, + 0.006465908605605364, + 0.008439844474196434, + 0.006060863379389048, + 0.005575988441705704, + 0.007799722719937563, + 0.00680515356361866, + 0.008795945905148983, + 0.006276676896959543, + 0.01661265827715397, + 0.008615418337285519, + 0.007406393066048622, + 0.0037499575410038233, + 0.004537641070783138, + 0.005096571519970894, + 0.00651358999311924, + 0.009860849007964134, + 0.015710795298218727, + 0.005962418392300606, + 0.012308257631957531, + 0.012300524860620499, + 0.03323003277182579, + 0.013880625367164612, + 0.014759672805666924, + 0.016597501933574677, + 0.017830874770879745, + 0.011469107121229172, + 0.02637092024087906, + 0.011863560415804386, + 0.012198892422020435, + 0.012199966236948967, + 0.010017012245953083, + 0.015969185158610344, + 0.00976747740060091, + 0.008393567055463791, + 0.012411640025675297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.003197202691808343 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.009780227206647396 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780227206647396, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780227206647396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11649848520755768, + 0.09215687215328217, + 0.12594173848628998, + 0.06846294552087784, + 0.0791885256767273, + 0.14550535380840302, + 0.09419944882392883, + 0.1036515161395073, + 0.0609746091067791, + 0.1134205311536789 + ], + "credit_map": [ + 0.016709832474589348, + 0.01927320845425129, + 0.01048985030502081, + 0.013498445972800255, + 0.01183860469609499, + 0.014016382396221161, + 0.015110938809812069, + 0.015561225824058056, + 0.00806108582764864, + 0.014981136657297611, + 0.008671391755342484, + 0.014923444017767906, + 0.012971065938472748, + 0.01189202256500721, + 0.014290712773799896, + 0.0063660116866230965, + 0.010150330141186714, + 0.017829909920692444, + 0.02307196520268917, + 0.016808629035949707, + 0.014306413941085339, + 0.012225518934428692, + 0.013492250815033913, + 0.018056720495224, + 0.0070305801928043365, + 0.0061806924641132355, + 0.011674242094159126, + 0.008358671329915524, + 0.005389704369008541, + 0.012935930863022804, + 0.007155132479965687, + 0.009737996384501457, + 0.007736769504845142, + 0.0058861980214715, + 0.010696899145841599, + 0.005678075365722179, + 0.005048161838203669, + 0.019943006336688995, + 0.005393158178776503, + 0.018806258216500282, + 0.011716162785887718, + 0.026424668729305267, + 0.01408920343965292, + 0.015177971683442593, + 0.030449412763118744, + 0.014126813039183617, + 0.01392557006329298, + 0.019595537334680557, + 0.007196631282567978, + 0.00950180646032095, + 0.01062303688377142, + 0.010506444610655308, + 0.008596686646342278, + 0.02704809233546257, + 0.00987318530678749, + 0.010853564366698265, + 0.006242130417376757, + 0.008497250266373158, + 0.011276123113930225, + 0.013804116286337376, + 0.019192347303032875, + 0.016719268634915352, + 0.009952224791049957, + 0.01796804741024971, + 0.005537147168070078, + 0.01151692122220993, + 0.007350834086537361, + 0.006446276791393757, + 0.008124987594783306, + 0.006900448352098465, + 0.007755179423838854, + 0.007342813070863485, + 0.013319665566086769, + 0.013906095176935196, + 0.017477761954069138, + 0.012507469393312931, + 0.015157799236476421, + 0.015202266164124012, + 0.010896001942455769, + 0.014953462406992912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.009780227206647396 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.05493060126900673 + ], + "opposite_logprob_tokens": [ + -2.9299306869506836 + ], + "logprob_reward": -0.05493060126900673, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493060126900673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10323688387870789, + 0.04734453931450844, + 0.2105368673801422, + 0.07418178766965866, + 0.0890187919139862, + 0.14839990437030792, + 0.05599036067724228, + 0.04416344687342644, + 0.08605080842971802, + 0.14107660949230194 + ], + "credit_map": [ + 0.012974494136869907, + 0.021055707708001137, + 0.007429931312799454, + 0.012121647596359253, + 0.013007830828428268, + 0.011755489744246006, + 0.013061003759503365, + 0.011830772273242474, + 0.005285119172185659, + 0.009219937026500702, + 0.004243120085448027, + 0.00667178351432085, + 0.00671276543289423, + 0.005278496537357569, + 0.0065175434574484825, + 0.003415769198909402, + 0.018090227618813515, + 0.02735893800854683, + 0.04181557521224022, + 0.025748364627361298, + 0.020457079634070396, + 0.021546101197600365, + 0.024026375263929367, + 0.03149419277906418, + 0.008328352123498917, + 0.007029624655842781, + 0.014031478203833103, + 0.009578370489180088, + 0.005743237212300301, + 0.011304332874715328, + 0.006445185746997595, + 0.011721203103661537, + 0.012256592512130737, + 0.008014069870114326, + 0.02298450842499733, + 0.008728922344744205, + 0.006125365849584341, + 0.015499968081712723, + 0.007906557992100716, + 0.007502809166908264, + 0.01308809407055378, + 0.015858476981520653, + 0.01192593201994896, + 0.012086882255971432, + 0.015062744729220867, + 0.036862947046756744, + 0.013682721182703972, + 0.02983209490776062, + 0.0052945539355278015, + 0.005172467790544033, + 0.007600081618875265, + 0.005745792295783758, + 0.00750823225826025, + 0.015078219585120678, + 0.005816399119794369, + 0.003774611745029688, + 0.0024635945446789265, + 0.0032620106358081102, + 0.005464216228574514, + 0.003617346053943038, + 0.005208401009440422, + 0.013802280649542809, + 0.0036096915137022734, + 0.006735904607921839, + 0.006032975390553474, + 0.01509203389286995, + 0.009641378186643124, + 0.007577619981020689, + 0.009085024707019329, + 0.008945907466113567, + 0.009885105304419994, + 0.019790755584836006, + 0.017569661140441895, + 0.016592981293797493, + 0.020800374448299408, + 0.016121579334139824, + 0.0214768648147583, + 0.016796039417386055, + 0.014931987971067429, + 0.016787102445960045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.05493060126900673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up depicts a mesmerizing vertical journey as a tiny bubble ascends through a crystal-clear, flaky mass of carbonated liquid contained in a sleek, minimalist glass. The scene is lit from the side by a soft, diffuse light source, casting clean, bright highlights along the curve of the glass and creating a luminous, almost ethereal quality. As the bubble begins its ascent from the bottom of the frame, the camera carefully tracks it; initially a perfect, shimmering microsphere, it subtly jiggles as it rises, its shape delicately deforming and vibrating under the pressure of the fluid. Its transparent membrane captures the light, refracting it into miniature rainbows with shimmering edges, while the clear liquid it passes through amplifies the softly blurred neutral background into a gentle bokeh. The bubble's steady, unwavering upward motion creates a tranquil and hypnotic atmosphere, focusing the viewer's full attention on this simple, beautiful phenomenon of physics and light as it moves from the bottom to the top of the frame.", + "seed": 4453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_2137_seed_4453_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": -0.013815929554402829, + "vlm_logprob": -0.013815929554402829, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.013815929554402829, + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.014477682299911976 + ], + "opposite_logprob_tokens": [ + -6.264477729797363 + ], + "logprob_reward": -0.014477682299911976, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.014477682299911976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06081397086381912, + 0.04169534891843796, + 0.03306020423769951, + 0.04441374912858009, + 0.05097747966647148, + 0.4442315697669983, + 0.02442649006843567, + 0.03449635207653046, + 0.04604227468371391, + 0.2198425531387329 + ], + "credit_map": [ + 0.006242766976356506, + 0.010575590655207634, + 0.007625133730471134, + 0.01072095800191164, + 0.008004296571016312, + 0.004277422558516264, + 0.007086793426424265, + 0.006281005218625069, + 0.003869464388117194, + 0.004220874048769474, + 0.0042978753335773945, + 0.008854439482092857, + 0.00465048011392355, + 0.002708662301301956, + 0.010027823969721794, + 0.0030657306779175997, + 0.003255438059568405, + 0.004032463766634464, + 0.0038956422358751297, + 0.004679047968238592, + 0.005157216917723417, + 0.0026622137520462275, + 0.006523653399199247, + 0.00285453163087368, + 0.004425625316798687, + 0.005695458967238665, + 0.006144126411527395, + 0.005846464540809393, + 0.006657048128545284, + 0.004880423657596111, + 0.006514774635434151, + 0.004249826539307833, + 0.006861678324639797, + 0.005830458365380764, + 0.005587058607488871, + 0.003682310925796628, + 0.003863383550196886, + 0.006675130687654018, + 0.0035537052899599075, + 0.014923756010830402, + 0.025780515745282173, + 0.032187823206186295, + 0.13442420959472656, + 0.02392398938536644, + 0.027521153911948204, + 0.14930805563926697, + 0.021730050444602966, + 0.029355760663747787, + 0.002138855867087841, + 0.004548829980194569, + 0.002188035286962986, + 0.005909166764467955, + 0.002720512682572007, + 0.002537202090024948, + 0.0029109905008226633, + 0.001472897594794631, + 0.0030791203025728464, + 0.004974629729986191, + 0.0034508283715695143, + 0.005252579692751169, + 0.0038850463461130857, + 0.0046623689122498035, + 0.0053484137170016766, + 0.003843365702778101, + 0.006429644301533699, + 0.005475527606904507, + 0.006834779400378466, + 0.005324167665094137, + 0.0059256767854094505, + 0.004374784417450428, + 0.008541051298379898, + 0.003136643208563328, + 0.022966040298342705, + 0.033303406089544296, + 0.020166683942079544, + 0.02628326788544655, + 0.0333990752696991, + 0.018692174926400185, + 0.037195801734924316, + 0.027836093679070473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.014477682299911976 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.011093014851212502 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093014851212502, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093014851212502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06622865051031113, + 0.04683714359998703, + 0.0816720724105835, + 0.023856161162257195, + 0.06120345741510391, + 0.13746599853038788, + 0.04755328223109245, + 0.08590655773878098, + 0.07233115285634995, + 0.37694552540779114 + ], + "credit_map": [ + 0.0072787110693752766, + 0.005565955303609371, + 0.01038382388651371, + 0.010693460702896118, + 0.007752038072794676, + 0.005271061323583126, + 0.009562356397509575, + 0.009721246547996998, + 0.003363358322530985, + 0.004994308575987816, + 0.004428746644407511, + 0.01074074488133192, + 0.0046616061590611935, + 0.005423765163868666, + 0.008594205603003502, + 0.004630409646779299, + 0.012934193946421146, + 0.00969018042087555, + 0.008420855738222599, + 0.007486370857805014, + 0.01246714685112238, + 0.009628502652049065, + 0.010867804288864136, + 0.01017702091485262, + 0.002613689983263612, + 0.0024924958124756813, + 0.0030640142504125834, + 0.002275792881846428, + 0.0035778190940618515, + 0.0023761431220918894, + 0.003277343697845936, + 0.004178863950073719, + 0.005026604980230331, + 0.010337632149457932, + 0.010115823708474636, + 0.007306658662855625, + 0.011462783440947533, + 0.003944487776607275, + 0.008128819987177849, + 0.004880645778030157, + 0.011033756658434868, + 0.012250843457877636, + 0.043747272342443466, + 0.007951685227453709, + 0.00974497850984335, + 0.037333954125642776, + 0.007791880983859301, + 0.007611625362187624, + 0.004984147381037474, + 0.005488818045705557, + 0.005689733196049929, + 0.012841688469052315, + 0.005303987301886082, + 0.005113341845571995, + 0.004403752740472555, + 0.0037278139498084784, + 0.007323962636291981, + 0.016711419448256493, + 0.008831239305436611, + 0.008474950678646564, + 0.011739418841898441, + 0.011838359758257866, + 0.014507007785141468, + 0.0064801983535289764, + 0.006567557342350483, + 0.007202569395303726, + 0.009019470773637295, + 0.011439570225775242, + 0.008607180789113045, + 0.007408163044601679, + 0.014125428162515163, + 0.007961216382682323, + 0.07595563679933548, + 0.05029436945915222, + 0.03620638698339462, + 0.031847912818193436, + 0.05318048968911171, + 0.03611239418387413, + 0.04745780676603317, + 0.04589052498340607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.011093014851212502 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09120336920022964, + 0.03952130675315857, + 0.04496141895651817, + 0.03315502032637596, + 0.057367876172065735, + 0.4252528250217438, + 0.07510323077440262, + 0.05403178557753563, + 0.06045662239193916, + 0.11894650757312775 + ], + "credit_map": [ + 0.009875088930130005, + 0.010013826191425323, + 0.014039348810911179, + 0.009923644363880157, + 0.011314916424453259, + 0.01087772287428379, + 0.011674902401864529, + 0.013483923859894276, + 0.0033900782000273466, + 0.004217811394482851, + 0.003654696047306061, + 0.009129378013312817, + 0.004219410475343466, + 0.0033581890165805817, + 0.008592329919338226, + 0.002959416015073657, + 0.005374818574637175, + 0.005235408432781696, + 0.00443313317373395, + 0.0047852471470832825, + 0.006974922958761454, + 0.005306681152433157, + 0.006026523187756538, + 0.006824690382927656, + 0.0037563350051641464, + 0.003679591231048107, + 0.003934399224817753, + 0.004502170253545046, + 0.0060836151242256165, + 0.003577547613531351, + 0.004586396738886833, + 0.003034967230632901, + 0.004955798853188753, + 0.008991583250463009, + 0.009568504989147186, + 0.008869227953255177, + 0.007261813618242741, + 0.005179679952561855, + 0.005891900043934584, + 0.006649371702224016, + 0.02456044778227806, + 0.03148133307695389, + 0.13536037504673004, + 0.022251924499869347, + 0.02452220395207405, + 0.14160531759262085, + 0.01865091547369957, + 0.02682034857571125, + 0.011254456825554371, + 0.011085227131843567, + 0.011455264873802662, + 0.009374232031404972, + 0.009395217522978783, + 0.007448866963386536, + 0.00858211051672697, + 0.00650785630568862, + 0.005234381649643183, + 0.00806630402803421, + 0.007302683312445879, + 0.005337620619684458, + 0.007771909236907959, + 0.007508648093789816, + 0.008432895876467228, + 0.004377350211143494, + 0.007119747810065746, + 0.006782723590731621, + 0.01077392976731062, + 0.006735077127814293, + 0.005727651994675398, + 0.009029140695929527, + 0.006963706109672785, + 0.007324653211981058, + 0.01060443464666605, + 0.025476718321442604, + 0.0077822161838412285, + 0.01463257148861885, + 0.023787464946508408, + 0.0100398613139987, + 0.018168967217206955, + 0.00845428928732872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.004639100283384323 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.012537471950054169 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537471950054169, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537471950054169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08004874736070633, + 0.025823289528489113, + 0.04362737387418747, + 0.055243976414203644, + 0.08822470158338547, + 0.5167393088340759, + 0.055967699736356735, + 0.05084073916077614, + 0.03504548594355583, + 0.048438556492328644 + ], + "credit_map": [ + 0.007384768687188625, + 0.013416820205748081, + 0.010628958232700825, + 0.010713719762861729, + 0.009175026789307594, + 0.007139085792005062, + 0.011912154965102673, + 0.009678220376372337, + 0.003058586036786437, + 0.0029285128694027662, + 0.00263753323815763, + 0.004614101257175207, + 0.00287646590732038, + 0.0030835021752864122, + 0.004215318243950605, + 0.0024092739913612604, + 0.007013032212853432, + 0.0050754244439303875, + 0.005466840695589781, + 0.004601853899657726, + 0.005227381829172373, + 0.006231669802218676, + 0.0049280282109975815, + 0.005083148833364248, + 0.007196999154984951, + 0.005151982419192791, + 0.006795378867536783, + 0.005320993717759848, + 0.005390532314777374, + 0.00616413401439786, + 0.006752656307071447, + 0.012471307069063187, + 0.010293962433934212, + 0.013743357732892036, + 0.016193291172385216, + 0.010335460305213928, + 0.009917841292917728, + 0.007829648442566395, + 0.007344346959143877, + 0.012566802091896534, + 0.029164060950279236, + 0.03780590742826462, + 0.16105997562408447, + 0.0275933425873518, + 0.03086749278008938, + 0.1733390986919403, + 0.024078890681266785, + 0.032830651849508286, + 0.007175929844379425, + 0.009373973123729229, + 0.008715211413800716, + 0.007172826211899519, + 0.008438635617494583, + 0.004997804760932922, + 0.006281254813075066, + 0.0038120748940855265, + 0.004781275987625122, + 0.008798860013484955, + 0.005409218370914459, + 0.007246475666761398, + 0.007134050130844116, + 0.0063688927330076694, + 0.00628929678350687, + 0.004812674596905708, + 0.003590691601857543, + 0.003983812406659126, + 0.00344870425760746, + 0.0038596412632614374, + 0.004185399506241083, + 0.006160399876534939, + 0.005505952052772045, + 0.004310886841267347, + 0.00698546227067709, + 0.00672110402956605, + 0.003326099133118987, + 0.005875986535102129, + 0.007997090928256512, + 0.0029932146426290274, + 0.010505234822630882, + 0.00403436878696084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.012537471950054169 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.026332378387451172 + ], + "opposite_logprob_tokens": [ + -3.651332378387451 + ], + "logprob_reward": -0.026332378387451172, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332378387451172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08537478744983673, + 0.044066816568374634, + 0.07041384279727936, + 0.037409283220767975, + 0.04633285850286484, + 0.40265321731567383, + 0.09136150777339935, + 0.10919170081615448, + 0.027500126510858536, + 0.08569583296775818 + ], + "credit_map": [ + 0.006706955377012491, + 0.014519525691866875, + 0.016659310087561607, + 0.012251988053321838, + 0.010683993808925152, + 0.005246966611593962, + 0.011359753087162971, + 0.007946286350488663, + 0.004815063904970884, + 0.0056966859847307205, + 0.0038496931083500385, + 0.00838569924235344, + 0.004832093138247728, + 0.004856274928897619, + 0.007875374518334866, + 0.003755931043997407, + 0.006646182853728533, + 0.010796280577778816, + 0.007614965550601482, + 0.009197495877742767, + 0.010631025768816471, + 0.006647674832493067, + 0.011561899445950985, + 0.007318318355828524, + 0.005329027771949768, + 0.0038536556530743837, + 0.004055759869515896, + 0.005361934192478657, + 0.005543180275708437, + 0.0041625346057116985, + 0.004669842775911093, + 0.0044333478435873985, + 0.00621790112927556, + 0.005562160164117813, + 0.005171456839889288, + 0.005907109938561916, + 0.00817756075412035, + 0.0035652138758450747, + 0.005181698128581047, + 0.006549758370965719, + 0.022807996720075607, + 0.028937673196196556, + 0.12210725992918015, + 0.021735988557338715, + 0.024264445528388023, + 0.13695235550403595, + 0.01901179365813732, + 0.02683570794761181, + 0.007818518206477165, + 0.012518838047981262, + 0.013434739783406258, + 0.01351857092231512, + 0.009538962505757809, + 0.01110264752060175, + 0.009006258100271225, + 0.01442297175526619, + 0.012164200656116009, + 0.020843306556344032, + 0.009257662110030651, + 0.01306708063930273, + 0.014916496351361275, + 0.013736719265580177, + 0.013188819400966167, + 0.012017426081001759, + 0.0027538607828319073, + 0.00278683309443295, + 0.0030736278276890516, + 0.0036258769687265158, + 0.0042372578755021095, + 0.0036618919111788273, + 0.004912428557872772, + 0.002448349492624402, + 0.010933181270956993, + 0.013250153511762619, + 0.009488273411989212, + 0.009566341526806355, + 0.010454636998474598, + 0.008102616295218468, + 0.013665037229657173, + 0.01023559458553791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.026332378387451172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up depicts a mesmerizing vertical journey as a tiny bubble ascends through a crystal-clear, flaky mass of carbonated liquid contained in a sleek, minimalist glass. The scene is lit from the side by a soft, diffuse light source, casting clean, bright highlights along the curve of the glass and creating a luminous, almost ethereal quality. As the bubble begins its ascent from the bottom of the frame, the camera carefully tracks it; initially a perfect, shimmering microsphere, it subtly jiggles as it rises, its shape delicately deforming and vibrating under the pressure of the fluid. Its transparent membrane captures the light, refracting it into miniature rainbows with shimmering edges, while the clear liquid it passes through amplifies the softly blurred neutral background into a gentle bokeh. The bubble's steady, unwavering upward motion creates a tranquil and hypnotic atmosphere, focusing the viewer's full attention on this simple, beautiful phenomenon of physics and light as it moves from the bottom to the top of the frame.", + "seed": 4454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_2137_seed_4454_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": -0.041552034113556145, + "vlm_logprob": -0.041552034113556145, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.041552034113556145, + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.010595260187983513 + ], + "opposite_logprob_tokens": [ + -5.385595321655273 + ], + "logprob_reward": -0.010595260187983513, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010595260187983513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3243314027786255, + 0.03462144732475281, + 0.07604095339775085, + 0.058128196746110916, + 0.16276036202907562, + 0.05773191526532173, + 0.05670074000954628, + 0.03137471154332161, + 0.08794105052947998, + 0.11036922037601471 + ], + "credit_map": [ + 0.024892808869481087, + 0.02638237178325653, + 0.08779766410589218, + 0.03239722549915314, + 0.02792816050350666, + 0.08680546283721924, + 0.019781338050961494, + 0.018346382305026054, + 0.0034638559445738792, + 0.0037608889397233725, + 0.004426825325936079, + 0.003914783243089914, + 0.00619759876281023, + 0.0039001780096441507, + 0.0050801802426576614, + 0.0038771401159465313, + 0.006033271085470915, + 0.004631411749869585, + 0.017478851601481438, + 0.006505289580672979, + 0.007940028794109821, + 0.018984422087669373, + 0.00566871790215373, + 0.008798959665000439, + 0.007254892960190773, + 0.006276467349380255, + 0.008698743768036366, + 0.005384764634072781, + 0.005985855124890804, + 0.007499010302126408, + 0.0071169668808579445, + 0.009911492466926575, + 0.015609592199325562, + 0.02114306204020977, + 0.02393304742872715, + 0.015498767606914043, + 0.025346729904413223, + 0.019146179780364037, + 0.026424983516335487, + 0.01565798558294773, + 0.0055795516818761826, + 0.005680720787495375, + 0.008014628663659096, + 0.004931527189910412, + 0.0075245145708322525, + 0.009892143309116364, + 0.009506010450422764, + 0.00660281628370285, + 0.005931832827627659, + 0.008034312166273594, + 0.00681251659989357, + 0.008705586194992065, + 0.009828699752688408, + 0.0037344596348702908, + 0.009109328500926495, + 0.004544003866612911, + 0.004265559371560812, + 0.0030330882873386145, + 0.005314315669238567, + 0.0037483510095626116, + 0.00401161378249526, + 0.004336605314165354, + 0.004021552391350269, + 0.0026436271145939827, + 0.011028806678950787, + 0.005355524364858866, + 0.014828234910964966, + 0.007778203580528498, + 0.00820283591747284, + 0.019424935802817345, + 0.009430976584553719, + 0.011891539208590984, + 0.01585542783141136, + 0.011359692551195621, + 0.013733302243053913, + 0.010260959155857563, + 0.009835964068770409, + 0.022904323413968086, + 0.01425242144614458, + 0.0121671287342906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.010595260187983513 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.07909199595451355 + ], + "opposite_logprob_tokens": [ + -2.579092025756836 + ], + "logprob_reward": -0.07909199595451355, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07909199595451355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2252766191959381, + 0.04478532820940018, + 0.08808073401451111, + 0.37675949931144714, + 0.11893751472234726, + 0.017730722203850746, + 0.01155711431056261, + 0.015284261666238308, + 0.018138831481337547, + 0.0834493339061737 + ], + "credit_map": [ + 0.02325359359383583, + 0.021913349628448486, + 0.03616045042872429, + 0.036514878273010254, + 0.03315897658467293, + 0.028838716447353363, + 0.024334609508514404, + 0.021102042868733406, + 0.004818638786673546, + 0.004804742522537708, + 0.006379483267664909, + 0.00488802045583725, + 0.006348560098558664, + 0.007872049696743488, + 0.005439725238829851, + 0.004234109539538622, + 0.007533169351518154, + 0.007342924829572439, + 0.020169580355286598, + 0.007984415628015995, + 0.010755294002592564, + 0.021451592445373535, + 0.006827926728874445, + 0.006015843711793423, + 0.049723777920007706, + 0.036576829850673676, + 0.06604591757059097, + 0.03846871480345726, + 0.04504852741956711, + 0.03929610922932625, + 0.06626512855291367, + 0.03533449396491051, + 0.012397125363349915, + 0.014631630852818489, + 0.018414009362459183, + 0.009876608848571777, + 0.014584220945835114, + 0.017466895282268524, + 0.016665711998939514, + 0.014901310205459595, + 0.0019822262693196535, + 0.0013671304332092404, + 0.0025595903862267733, + 0.0013214024947956204, + 0.0017601626459509134, + 0.0036367280408740044, + 0.002343288157135248, + 0.0027601944748312235, + 0.0015829583862796426, + 0.0011091529158875346, + 0.0014290588442236185, + 0.0013167475117370486, + 0.0009992594132199883, + 0.002247123746201396, + 0.001107610180042684, + 0.0017652040114626288, + 0.0021316115744411945, + 0.0023475608322769403, + 0.0017886898713186383, + 0.0029270704835653305, + 0.0014535695081576705, + 0.0019792462699115276, + 0.0011328301625326276, + 0.0015236835461109877, + 0.002923043444752693, + 0.0010464794468134642, + 0.0029000521171838045, + 0.0010761176235973835, + 0.0017288895323872566, + 0.003706888761371374, + 0.0018274376634508371, + 0.002929922891780734, + 0.014618118293583393, + 0.009354270063340664, + 0.009898439981043339, + 0.008184392936527729, + 0.005903101060539484, + 0.016292300075292587, + 0.012009740807116032, + 0.007188967429101467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.07909199595451355 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18895921111106873, + 0.04607173055410385, + 0.07362107932567596, + 0.11861073970794678, + 0.3977697491645813, + 0.04025135189294815, + 0.026703689247369766, + 0.019927978515625, + 0.03361302614212036, + 0.054471455514431 + ], + "credit_map": [ + 0.01958022452890873, + 0.023141667246818542, + 0.027130506932735443, + 0.028158709406852722, + 0.022249441593885422, + 0.0314384289085865, + 0.02596164122223854, + 0.011298568919301033, + 0.004297949839383364, + 0.004171490669250488, + 0.006467647384852171, + 0.004675670526921749, + 0.007236431818455458, + 0.008683315478265285, + 0.0068824621848762035, + 0.0036567645147442818, + 0.0063890619203448296, + 0.005932318512350321, + 0.01281742099672556, + 0.007285566069185734, + 0.006416831165552139, + 0.018359709531068802, + 0.005059794057160616, + 0.011360371485352516, + 0.01545457448810339, + 0.011040523648262024, + 0.018185513094067574, + 0.012226229533553123, + 0.012592621147632599, + 0.011869929730892181, + 0.02056615613400936, + 0.016675185412168503, + 0.03828649967908859, + 0.048402175307273865, + 0.0491299107670784, + 0.039918117225170135, + 0.07167535275220871, + 0.042857736349105835, + 0.07383690774440765, + 0.033663053065538406, + 0.003944998607039452, + 0.005948815494775772, + 0.005001877900213003, + 0.0038419715128839016, + 0.005595343187451363, + 0.004491297993808985, + 0.006857638712972403, + 0.004569406621158123, + 0.003729293355718255, + 0.0022388894576579332, + 0.005053199362009764, + 0.004259475506842136, + 0.0031136558391153812, + 0.0029489207081496716, + 0.0031113498844206333, + 0.00224890629760921, + 0.0028535339515656233, + 0.0018410407938063145, + 0.003318630624562502, + 0.002243742346763611, + 0.002977146999910474, + 0.0028060534968972206, + 0.002177604241296649, + 0.0017102262936532497, + 0.004176266957074404, + 0.002834981307387352, + 0.00430833175778389, + 0.0038066054694354534, + 0.004170795902609825, + 0.0041421204805374146, + 0.006452016532421112, + 0.0037219091318547726, + 0.010337108746170998, + 0.005764491390436888, + 0.006931565701961517, + 0.005044145043939352, + 0.004162870813161135, + 0.010839962400496006, + 0.006566355004906654, + 0.004824953619390726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005256166215986013 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18431887030601501, + 0.07264478504657745, + 0.14234323799610138, + 0.18521511554718018, + 0.1782470941543579, + 0.02473251335322857, + 0.04570474475622177, + 0.03354932367801666, + 0.09194224327802658, + 0.04130212217569351 + ], + "credit_map": [ + 0.02179584465920925, + 0.026050914078950882, + 0.02020857110619545, + 0.03010820783674717, + 0.02490817941725254, + 0.019961964339017868, + 0.02917023003101349, + 0.012114960700273514, + 0.008318571373820305, + 0.006913487333804369, + 0.0107346186414361, + 0.008137437514960766, + 0.0099351080134511, + 0.010180702432990074, + 0.010489467531442642, + 0.00793539173901081, + 0.013541902415454388, + 0.01355996634811163, + 0.02043968439102173, + 0.013514382764697075, + 0.01621316187083721, + 0.03244279697537422, + 0.009129820391535759, + 0.02350151166319847, + 0.02746501751244068, + 0.02050074003636837, + 0.03789407014846802, + 0.017331566661596298, + 0.021417899057269096, + 0.016521098092198372, + 0.030257178470492363, + 0.013827549293637276, + 0.019095784053206444, + 0.01925016939640045, + 0.023943958804011345, + 0.013827163726091385, + 0.029732339084148407, + 0.022613734006881714, + 0.026281483471393585, + 0.023502454161643982, + 0.004239010624587536, + 0.0029012029990553856, + 0.0036244308575987816, + 0.002118234522640705, + 0.003803635947406292, + 0.0028177069034427404, + 0.0027730511501431465, + 0.002455240348353982, + 0.0046159070916473866, + 0.005930780898779631, + 0.0056901308707892895, + 0.007630515843629837, + 0.007067815400660038, + 0.004104039631783962, + 0.006224191281944513, + 0.004441367462277412, + 0.004138903226703405, + 0.004119192715734243, + 0.005062547978013754, + 0.003944905940443277, + 0.0058752805925905704, + 0.003534568240866065, + 0.003827395848929882, + 0.0030465286690741777, + 0.010631578974425793, + 0.008718162775039673, + 0.011461073532700539, + 0.010167884640395641, + 0.013850816525518894, + 0.010768544860184193, + 0.018265239894390106, + 0.008078938350081444, + 0.004241702146828175, + 0.005259320139884949, + 0.0047627706080675125, + 0.005149008706212044, + 0.004562695045024157, + 0.005079556722193956, + 0.007949409075081348, + 0.00429765647277236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.01253888476639986 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.1002778634428978 + ], + "opposite_logprob_tokens": [ + -2.350277900695801 + ], + "logprob_reward": -0.1002778634428978, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002778634428978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11632125824689865, + 0.05945833399891853, + 0.14185765385627747, + 0.04914368316531181, + 0.1992148756980896, + 0.20718613266944885, + 0.059737272560596466, + 0.02291438728570938, + 0.049471039324998856, + 0.09469544142484665 + ], + "credit_map": [ + 0.012322794646024704, + 0.011992527171969414, + 0.014325940050184727, + 0.021024053916335106, + 0.020330922678112984, + 0.013591334223747253, + 0.014164327643811703, + 0.008569354191422462, + 0.0060266475193202496, + 0.006012279074639082, + 0.006621440406888723, + 0.007885873317718506, + 0.007296144962310791, + 0.00500886607915163, + 0.015081429854035378, + 0.005525648593902588, + 0.013743430376052856, + 0.01279482338577509, + 0.019532056525349617, + 0.014996238052845001, + 0.01696876809000969, + 0.02970801666378975, + 0.01094739604741335, + 0.023166915401816368, + 0.006071491166949272, + 0.00481253769248724, + 0.00780260656028986, + 0.005886654369533062, + 0.006761317141354084, + 0.004288834985345602, + 0.009088604710996151, + 0.004431632347404957, + 0.024394366890192032, + 0.023155400529503822, + 0.027177786454558372, + 0.016836171969771385, + 0.027380380779504776, + 0.023342235013842583, + 0.03014601580798626, + 0.026782497763633728, + 0.02116948552429676, + 0.02198905497789383, + 0.0251334086060524, + 0.016749219968914986, + 0.030081097036600113, + 0.03866928443312645, + 0.029793664813041687, + 0.02360089123249054, + 0.007468605414032936, + 0.006279064808040857, + 0.007571776397526264, + 0.007398459129035473, + 0.007959778420627117, + 0.008662989363074303, + 0.006321370601654053, + 0.008075222373008728, + 0.0026167836040258408, + 0.003720863489434123, + 0.003023802302777767, + 0.0033781221136450768, + 0.0033979376312345266, + 0.0026590670458972454, + 0.0021207842510193586, + 0.001997023355215788, + 0.006132034584879875, + 0.0033035397063940763, + 0.007441281341016293, + 0.004089341964572668, + 0.005791252013295889, + 0.010257241316139698, + 0.006608231924474239, + 0.005848110653460026, + 0.011689425446093082, + 0.009010732173919678, + 0.011273444630205631, + 0.008322236128151417, + 0.008287577889859676, + 0.019962197169661522, + 0.013047508895397186, + 0.013102307915687561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.1002778634428978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up depicts a mesmerizing vertical journey as a tiny bubble ascends through a crystal-clear, flaky mass of carbonated liquid contained in a sleek, minimalist glass. The scene is lit from the side by a soft, diffuse light source, casting clean, bright highlights along the curve of the glass and creating a luminous, almost ethereal quality. As the bubble begins its ascent from the bottom of the frame, the camera carefully tracks it; initially a perfect, shimmering microsphere, it subtly jiggles as it rises, its shape delicately deforming and vibrating under the pressure of the fluid. Its transparent membrane captures the light, refracting it into miniature rainbows with shimmering edges, while the clear liquid it passes through amplifies the softly blurred neutral background into a gentle bokeh. The bubble's steady, unwavering upward motion creates a tranquil and hypnotic atmosphere, focusing the viewer's full attention on this simple, beautiful phenomenon of physics and light as it moves from the bottom to the top of the frame.", + "seed": 4455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_15/index_2137_seed_4455_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": -0.02428892981261015, + "vlm_logprob": -0.02428892981261015, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02428892981261015, + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.016651811078190804 + ], + "opposite_logprob_tokens": [ + -4.391651630401611 + ], + "logprob_reward": -0.016651811078190804, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016651811078190804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18613219261169434, + 0.12431342154741287, + 0.07213620841503143, + 0.15072785317897797, + 0.07874327152967453, + 0.15245658159255981, + 0.0333993062376976, + 0.025406373664736748, + 0.05243103578686714, + 0.12425374984741211 + ], + "credit_map": [ + 0.01965811289846897, + 0.018997926265001297, + 0.035226721316576004, + 0.01602076180279255, + 0.021369963884353638, + 0.0406591072678566, + 0.016800563782453537, + 0.017399048432707787, + 0.015681594610214233, + 0.013015447184443474, + 0.023695174604654312, + 0.013683263212442398, + 0.011932868510484695, + 0.023984704166650772, + 0.012010344304144382, + 0.010310019366443157, + 0.007607020437717438, + 0.006662553641945124, + 0.007339611183851957, + 0.009079374372959137, + 0.008782081305980682, + 0.006598303560167551, + 0.013819126412272453, + 0.012248139828443527, + 0.01777813956141472, + 0.01258072443306446, + 0.01746556907892227, + 0.018542692065238953, + 0.03440985083580017, + 0.01627357117831707, + 0.020025210455060005, + 0.013652111403644085, + 0.010341753251850605, + 0.007435751147568226, + 0.009641090407967567, + 0.0075436183251440525, + 0.017574336379766464, + 0.00830153189599514, + 0.010444004088640213, + 0.007461184170097113, + 0.014114039950072765, + 0.013196869753301144, + 0.02158285863697529, + 0.021468373015522957, + 0.03230711445212364, + 0.013117494061589241, + 0.024630190804600716, + 0.012039651162922382, + 0.0033499011769890785, + 0.003539959667250514, + 0.005228764843195677, + 0.0038621078711003065, + 0.006372720003128052, + 0.0031600079964846373, + 0.004626638256013393, + 0.0032592054922133684, + 0.0032736784778535366, + 0.002311831573024392, + 0.003324624616652727, + 0.0024195280857384205, + 0.00318133900873363, + 0.004085127729922533, + 0.0029425963293761015, + 0.003867649007588625, + 0.006021261215209961, + 0.0030277473852038383, + 0.012554651126265526, + 0.003385040210559964, + 0.005264663603156805, + 0.011678741313517094, + 0.004080932587385178, + 0.006417999044060707, + 0.013499056920409203, + 0.008217379450798035, + 0.025774361565709114, + 0.009504683315753937, + 0.018040362745523453, + 0.010210657492280006, + 0.02475465089082718, + 0.01425259094685316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.016651811078190804 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.011090303771197796 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090303771197796, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090303771197796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1894303858280182, + 0.36156776547431946, + 0.059375301003456116, + 0.051036469638347626, + 0.07154075056314468, + 0.04746397212147713, + 0.047858793288469315, + 0.04601098224520683, + 0.0387260727584362, + 0.08698950707912445 + ], + "credit_map": [ + 0.019265467301011086, + 0.022830402478575706, + 0.03190155699849129, + 0.02366809919476509, + 0.02096727304160595, + 0.025853630155324936, + 0.02528378926217556, + 0.019660163670778275, + 0.03035399504005909, + 0.018866954371333122, + 0.11594360321760178, + 0.024162787944078445, + 0.016595179215073586, + 0.11040907353162766, + 0.02021877095103264, + 0.025017371401190758, + 0.005701620131731033, + 0.005856000352650881, + 0.00825783796608448, + 0.007712505757808685, + 0.0065539125353097916, + 0.00578988716006279, + 0.011324127204716206, + 0.008179412223398685, + 0.0062228357419371605, + 0.0041850777342915535, + 0.008131907321512699, + 0.007157956715673208, + 0.006651116069406271, + 0.00646174605935812, + 0.006910030264407396, + 0.005315800197422504, + 0.007976838387548923, + 0.0066417185589671135, + 0.008541649207472801, + 0.007886105217039585, + 0.012623037211596966, + 0.008763939142227173, + 0.011017955839633942, + 0.008089512586593628, + 0.005108759272843599, + 0.004156075417995453, + 0.0064905607141554356, + 0.007783577777445316, + 0.006488082930445671, + 0.0044659399427473545, + 0.007152271922677755, + 0.0058187018148601055, + 0.005096155684441328, + 0.003758228849619627, + 0.006539232097566128, + 0.005626017693430185, + 0.007941863499581814, + 0.006245569325983524, + 0.007504247594624758, + 0.005147475283592939, + 0.0058512212708592415, + 0.004618009086698294, + 0.006233141291886568, + 0.003917437046766281, + 0.008225916884839535, + 0.006079911720007658, + 0.004957882221788168, + 0.006127461325377226, + 0.006030488759279251, + 0.002901615109294653, + 0.006469260435551405, + 0.0035620764829218388, + 0.005135771818459034, + 0.005123431794345379, + 0.004095368552953005, + 0.005408060271292925, + 0.008932611905038357, + 0.008510924875736237, + 0.014190453104674816, + 0.013454711996018887, + 0.013364923186600208, + 0.007893551141023636, + 0.012313107028603554, + 0.008329224772751331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.011090303771197796 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.0076340362429618835 + ], + "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": [ + 0.1520746946334839, + 0.19797930121421814, + 0.046804144978523254, + 0.06557981669902802, + 0.10501786321401596, + 0.18077987432479858, + 0.043361786752939224, + 0.05726051330566406, + 0.0552937313914299, + 0.09584834426641464 + ], + "credit_map": [ + 0.015168692916631699, + 0.019824430346488953, + 0.020892878994345665, + 0.016880134120583534, + 0.02285960502922535, + 0.018402978777885437, + 0.017091555520892143, + 0.020954404026269913, + 0.01385426614433527, + 0.011859806254506111, + 0.06295223534107208, + 0.014524005353450775, + 0.010083758272230625, + 0.056866735219955444, + 0.014375424943864346, + 0.01346303429454565, + 0.005632954649627209, + 0.0031404029577970505, + 0.005122616421431303, + 0.005178909283131361, + 0.005286187399178743, + 0.0052632614970207214, + 0.00794786773622036, + 0.009231936186552048, + 0.007351630367338657, + 0.005012561101466417, + 0.008002344518899918, + 0.009491486474871635, + 0.012180589139461517, + 0.0086526395753026, + 0.008033272810280323, + 0.006855288986116648, + 0.01307976059615612, + 0.010160764679312706, + 0.014571317471563816, + 0.010139429941773415, + 0.022831909358501434, + 0.009987778030335903, + 0.014321689493954182, + 0.00992520060390234, + 0.016924528405070305, + 0.013214796781539917, + 0.02138596959412098, + 0.024623945355415344, + 0.03656645864248276, + 0.016457578167319298, + 0.037919025868177414, + 0.01368755754083395, + 0.004088459070771933, + 0.004892949014902115, + 0.0061671859584748745, + 0.005160587839782238, + 0.006037744227796793, + 0.004184332676231861, + 0.008938610553741455, + 0.0038919115904718637, + 0.005338430870324373, + 0.0037630470469594, + 0.011824581772089005, + 0.003919281996786594, + 0.010753598064184189, + 0.01219878438860178, + 0.005657373461872339, + 0.00380541174672544, + 0.005216280929744244, + 0.004077840130776167, + 0.013840433210134506, + 0.004339909180998802, + 0.005956193432211876, + 0.011613892391324043, + 0.005097534041851759, + 0.005151642486453056, + 0.00881771370768547, + 0.009056752547621727, + 0.011175509542226791, + 0.010574228130280972, + 0.015360034070909023, + 0.009607894346117973, + 0.02004600502550602, + 0.011210194788873196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1815311759710312, + 0.2560868263244629, + 0.04404940456151962, + 0.07064611464738846, + 0.08567792922258377, + 0.1146475076675415, + 0.04181348532438278, + 0.0392819307744503, + 0.03758828714489937, + 0.12867726385593414 + ], + "credit_map": [ + 0.017880044877529144, + 0.017556747421622276, + 0.025426097214221954, + 0.012747442349791527, + 0.025538770481944084, + 0.027030935510993004, + 0.018262730911374092, + 0.037088412791490555, + 0.01896739937365055, + 0.01174798235297203, + 0.0909961611032486, + 0.01548093743622303, + 0.012995772063732147, + 0.07948629558086395, + 0.011685745790600777, + 0.014726550318300724, + 0.005436848383396864, + 0.0033577154390513897, + 0.008301510475575924, + 0.005923996679484844, + 0.002925756387412548, + 0.005751584656536579, + 0.006917896680533886, + 0.005434098187834024, + 0.008337108418345451, + 0.005216837394982576, + 0.00795577373355627, + 0.011789344251155853, + 0.01327548362314701, + 0.006933214608579874, + 0.010613043792545795, + 0.0065253060311079025, + 0.010833675041794777, + 0.007540796883404255, + 0.01096202153712511, + 0.009421571157872677, + 0.018237583339214325, + 0.008221388794481754, + 0.0127233462408185, + 0.007737543899565935, + 0.013336812146008015, + 0.007504373323172331, + 0.012023038230836391, + 0.015661733224987984, + 0.018952352926135063, + 0.011479112319648266, + 0.02490067109465599, + 0.010789408348500729, + 0.004609229508787394, + 0.004197436850517988, + 0.005917830858379602, + 0.004486261401325464, + 0.00842104759067297, + 0.004443197976797819, + 0.006204572040587664, + 0.0035339100286364555, + 0.005802810657769442, + 0.0029973406344652176, + 0.006266231648623943, + 0.002257782267406583, + 0.005278835538774729, + 0.007706582546234131, + 0.003369291080161929, + 0.005603056401014328, + 0.003565547987818718, + 0.002214741427451372, + 0.008711498230695724, + 0.002900694729760289, + 0.004615887068212032, + 0.007909197360277176, + 0.00417099054902792, + 0.00349973002448678, + 0.014891131781041622, + 0.010697473771870136, + 0.01693427748978138, + 0.01171020045876503, + 0.02145443670451641, + 0.011819656938314438, + 0.022364087402820587, + 0.01880599930882454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.01606813259422779 + }, + { + "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 tiny bubble ascending through a carbonated liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the liquid contained within a sleek, minimalist glass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bubble exhibit a jiggling, deforming, and vibrating motion as it rises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bubble refract light into miniature rainbows with shimmering edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track the bubble's upward motion from the bottom to the top of the 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": [ + -0.07000036537647247 + ], + "opposite_logprob_tokens": [ + -2.695000410079956 + ], + "logprob_reward": -0.07000036537647247, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000036537647247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35960283875465393, + 0.2110493928194046, + 0.05569906160235405, + 0.05662195384502411, + 0.031969282776117325, + 0.04537014663219452, + 0.06724804639816284, + 0.052959613502025604, + 0.03457139804959297, + 0.08490833640098572 + ], + "credit_map": [ + 0.031265802681446075, + 0.032279081642627716, + 0.0806446522474289, + 0.04003321751952171, + 0.03760385513305664, + 0.05318539962172508, + 0.04295600578188896, + 0.041634805500507355, + 0.0139321219176054, + 0.01030697301030159, + 0.06821813434362411, + 0.01434503123164177, + 0.009625599719583988, + 0.06567718088626862, + 0.013595487922430038, + 0.015348868444561958, + 0.006942107807844877, + 0.004506798926740885, + 0.009165467694401741, + 0.006945858709514141, + 0.006915481295436621, + 0.006514217704534531, + 0.007975099608302116, + 0.006734027527272701, + 0.007111186161637306, + 0.004979001823812723, + 0.008402331732213497, + 0.006863017566502094, + 0.009596683084964752, + 0.006738500203937292, + 0.007473306730389595, + 0.005457929335534573, + 0.00366048701107502, + 0.002550621284171939, + 0.004568507894873619, + 0.00507833156734705, + 0.004350820556282997, + 0.0031457345467060804, + 0.005034707020968199, + 0.003580074291676283, + 0.006078224163502455, + 0.0036152058746665716, + 0.006773675326257944, + 0.005501612089574337, + 0.0045065609738230705, + 0.005366711877286434, + 0.00676776934415102, + 0.006760388147085905, + 0.006506488658487797, + 0.006498520262539387, + 0.011801577173173428, + 0.007139049936085939, + 0.012386969290673733, + 0.006056175101548433, + 0.010618655011057854, + 0.0062406063079833984, + 0.006051427684724331, + 0.004469059407711029, + 0.006851616781204939, + 0.0041737025603652, + 0.011840606108307838, + 0.009078999981284142, + 0.006179431919008493, + 0.004314773250371218, + 0.0035499343648552895, + 0.002875284058973193, + 0.00746571458876133, + 0.003067182609811425, + 0.004362011328339577, + 0.006702662445604801, + 0.0032510750461369753, + 0.0032975315116345882, + 0.00869441032409668, + 0.009400146082043648, + 0.010995142161846161, + 0.007927405647933483, + 0.014325126074254513, + 0.007085510995239019, + 0.01521083153784275, + 0.011269758455455303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.07000036537647247 + } + ], + "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/0000017/group_2/index_tensor([7006, 7006])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_tensor([7006, 7006])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ac965a2d148f09c3623ca63d8a0d2959a98d3805 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_tensor([7006, 7006])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic sequence imbued with dynamism and tension, composed of a series of rapid, violent edits, opens with a turbulent long shot of a weathered, dark wooden fishing boat bobbing violently on a turbulent, gray-black sea. Under a gloomy, stormy sky, waves crash against the boat like mountains. Then, through abrupt jump cuts, the scene switches back and forth between two striking close-ups: one is an extreme close-up of thorny, dark green vines, as thick as pythons, tightly wrapped around the wet side and mast, nearly strangling the boat. The vines, covered in sharp black thorns, are stretched straight under the immense tension, and even tiny splinters can be seen breaking from the squeezed planking. The other, captured from an extremely low angle, captures a massive, white-foaming wave crashing against the side of the boat, instantly exploding into a cloud of water droplets and mist, each droplet appearing crystal clear and cold in the dim light. The entire sequence is gloomy in color, dominated by deep blue, dark green, and the dark brown of wet wood, with only the white foam of the waves providing a harsh contrast. The rapid editing rhythm combined with the helpless tossing of the hull creates a tense atmosphere of struggle and despair, surrounded by the violence of nature.", + "seed": 22036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22036_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": -0.021035246597602964, + "vlm_logprob": -0.021035246597602964, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.021035246597602964, + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032097600400447845 + ], + "opposite_logprob_tokens": [ + -4.157097816467285 + ], + "logprob_reward": -0.032097600400447845, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.032097600400447845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14948469400405884, + 0.12811794877052307, + 0.07422240078449249, + 0.06263261288404465, + 0.06071634963154793, + 0.2756198048591614, + 0.05591810122132301, + 0.05358577519655228, + 0.06265196204185486, + 0.07705040276050568 + ], + "credit_map": [ + 0.020017333328723907, + 0.015700049698352814, + 0.02478635124862194, + 0.015905803069472313, + 0.015647107735276222, + 0.02170218899846077, + 0.014436090365052223, + 0.021289754658937454, + 0.01197702344506979, + 0.019283028319478035, + 0.016804296523332596, + 0.01917753554880619, + 0.014195485971868038, + 0.010876339860260487, + 0.018589956685900688, + 0.01721426658332348, + 0.0054459115490317345, + 0.004060168284922838, + 0.013971545733511448, + 0.010883908718824387, + 0.006594550795853138, + 0.006580971647053957, + 0.015610010363161564, + 0.011075322516262531, + 0.004038542043417692, + 0.00504571245983243, + 0.01094865519553423, + 0.009548169560730457, + 0.007483154535293579, + 0.008258300833404064, + 0.00886158924549818, + 0.008448483422398567, + 0.0034410813823342323, + 0.0050945766270160675, + 0.014556551352143288, + 0.009895688854157925, + 0.008548244833946228, + 0.00448955874890089, + 0.008699169382452965, + 0.005991474259644747, + 0.009273265488445759, + 0.032990463078022, + 0.018943630158901215, + 0.10109375417232513, + 0.07172743231058121, + 0.012254983186721802, + 0.01794753596186638, + 0.0113886883482337, + 0.006188360974192619, + 0.011973495595157146, + 0.0061077214777469635, + 0.00908060371875763, + 0.006259204354137182, + 0.004642829764634371, + 0.00639725849032402, + 0.0052686212584376335, + 0.0052169677801430225, + 0.007547017652541399, + 0.006287754513323307, + 0.004948736634105444, + 0.007800016086548567, + 0.006656697951257229, + 0.006193758919835091, + 0.008934815414249897, + 0.007096258923411369, + 0.009324394166469574, + 0.007689198479056358, + 0.009117976762354374, + 0.006715482100844383, + 0.00847703218460083, + 0.007950431667268276, + 0.00628118310123682, + 0.0071340687572956085, + 0.005402879789471626, + 0.015466536395251751, + 0.010778303258121014, + 0.00687287375330925, + 0.007922245189547539, + 0.012181771919131279, + 0.011291717179119587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.032097600400447845 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016117636114358902 + ], + "opposite_logprob_tokens": [ + -4.141117572784424 + ], + "logprob_reward": -0.016117636114358902, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016117636114358902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13745690882205963, + 0.07422932982444763, + 0.11860645562410355, + 0.078999362885952, + 0.10561848431825638, + 0.18611693382263184, + 0.10597001016139984, + 0.05223065987229347, + 0.04835148900747299, + 0.09242041409015656 + ], + "credit_map": [ + 0.02411467395722866, + 0.018093029037117958, + 0.024782706052064896, + 0.012713632546365261, + 0.011502744629979134, + 0.02297736518085003, + 0.011580582708120346, + 0.011692163534462452, + 0.006516479421406984, + 0.007139049936085939, + 0.008014797233045101, + 0.009159969165921211, + 0.011788560077548027, + 0.005996544845402241, + 0.013523119501769543, + 0.012090817093849182, + 0.008825370110571384, + 0.006692370865494013, + 0.016934681683778763, + 0.013723579235374928, + 0.009176432155072689, + 0.015424112789332867, + 0.02635277435183525, + 0.02147712931036949, + 0.005086175166070461, + 0.006321114953607321, + 0.016237948089838028, + 0.014756311662495136, + 0.005111551843583584, + 0.008929602801799774, + 0.012490229681134224, + 0.010066431015729904, + 0.008422593586146832, + 0.010237405076622963, + 0.02499813586473465, + 0.016736099496483803, + 0.01398780383169651, + 0.007727683987468481, + 0.014031247235834599, + 0.009477517567574978, + 0.009629516862332821, + 0.039354462176561356, + 0.022165294736623764, + 0.03425566107034683, + 0.041171759366989136, + 0.011649337597191334, + 0.013746731914579868, + 0.01414416078478098, + 0.009819062426686287, + 0.025696920230984688, + 0.01414396334439516, + 0.011278166435658932, + 0.015281092375516891, + 0.010874717496335506, + 0.007265560794621706, + 0.011610525660216808, + 0.004761223681271076, + 0.008968917652964592, + 0.006279694382101297, + 0.004795507062226534, + 0.00810184981673956, + 0.006048950366675854, + 0.0068011898547410965, + 0.006473328452557325, + 0.004938532132655382, + 0.006467817351222038, + 0.006171219050884247, + 0.005683206021785736, + 0.006518179085105658, + 0.007391540799289942, + 0.006286365445703268, + 0.004894629120826721, + 0.006172847468405962, + 0.005140822846442461, + 0.016064224764704704, + 0.01428762823343277, + 0.011471646837890148, + 0.005983682349324226, + 0.022602448239922523, + 0.010697110556066036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.016117636114358902 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048705194145441055 + ], + "opposite_logprob_tokens": [ + -3.0487051010131836 + ], + "logprob_reward": -0.048705194145441055, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048705194145441055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884719118475914, + 0.14082573354244232, + 0.16241566836833954, + 0.09817788749933243, + 0.1599520742893219, + 0.13846661150455475, + 0.06858275830745697, + 0.052128054201602936, + 0.041414350271224976, + 0.0495649017393589 + ], + "credit_map": [ + 0.016230372712016106, + 0.009393041022121906, + 0.01434002909809351, + 0.007567510474473238, + 0.008889885619282722, + 0.014021155424416065, + 0.008563423529267311, + 0.009466501884162426, + 0.012645100243389606, + 0.01826341450214386, + 0.01564209908246994, + 0.012572885490953922, + 0.030500732362270355, + 0.014797004871070385, + 0.022081369534134865, + 0.014323131181299686, + 0.01133474800735712, + 0.017653826624155045, + 0.02222355268895626, + 0.029491495341062546, + 0.011083792895078659, + 0.01961638405919075, + 0.02487107738852501, + 0.02614080160856247, + 0.006499245297163725, + 0.006737713236361742, + 0.0219215527176857, + 0.01145294588059187, + 0.009021679870784283, + 0.01351870596408844, + 0.011416989378631115, + 0.017609059810638428, + 0.007444147486239672, + 0.01632506586611271, + 0.029016204178333282, + 0.03676838055253029, + 0.025067787617444992, + 0.010920862667262554, + 0.025352496653795242, + 0.009057139977812767, + 0.009226949885487556, + 0.03432631120085716, + 0.01563781313598156, + 0.020402900874614716, + 0.030074061825871468, + 0.006411446258425713, + 0.012706107459962368, + 0.009681017138063908, + 0.008293364197015762, + 0.016122737899422646, + 0.006668287795037031, + 0.009278565645217896, + 0.010229608044028282, + 0.0051047164015471935, + 0.0060683852061629295, + 0.006817101500928402, + 0.005154619459062815, + 0.007352638058364391, + 0.005895616021007299, + 0.0043164510279893875, + 0.008189712651073933, + 0.006474871188402176, + 0.006032364442944527, + 0.008711788803339005, + 0.0035493597388267517, + 0.00518024479970336, + 0.005695254076272249, + 0.005757316946983337, + 0.004017987754195929, + 0.004669602029025555, + 0.00679908599704504, + 0.005745501257479191, + 0.004500035662204027, + 0.005337209906429052, + 0.0068268124014139175, + 0.007185439579188824, + 0.005021721124649048, + 0.003541091689839959, + 0.011378467082977295, + 0.005774125922471285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.048705194145441055 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001522573409602046 + ], + "opposite_logprob_tokens": [ + -6.501522541046143 + ], + "logprob_reward": -0.001522573409602046, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001522573409602046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2128704935312271, + 0.14841189980506897, + 0.06581331044435501, + 0.05632902681827545, + 0.08545146882534027, + 0.11390049010515213, + 0.06395789980888367, + 0.11077114939689636, + 0.06728850305080414, + 0.07520575076341629 + ], + "credit_map": [ + 0.03627815097570419, + 0.024470919743180275, + 0.032846599817276, + 0.022403938695788383, + 0.018558712676167488, + 0.03750346973538399, + 0.018499048426747322, + 0.022309642285108566, + 0.013856622390449047, + 0.015094106085598469, + 0.01633630134165287, + 0.013416673056781292, + 0.029115622863173485, + 0.01571807824075222, + 0.026092059910297394, + 0.01878240332007408, + 0.005503444001078606, + 0.007695378735661507, + 0.009079569950699806, + 0.009486514143645763, + 0.006321708671748638, + 0.009530877694487572, + 0.009940547868609428, + 0.00825525913387537, + 0.003510805079713464, + 0.0053233192302286625, + 0.011611363850533962, + 0.009841930121183395, + 0.005983234383165836, + 0.005959708243608475, + 0.005909104831516743, + 0.008189554326236248, + 0.0061754765920341015, + 0.009555323049426079, + 0.0161734726279974, + 0.014984263107180595, + 0.013512136414647102, + 0.006048617884516716, + 0.012556902132928371, + 0.0064452714286744595, + 0.0065839579328894615, + 0.019900277256965637, + 0.010363921523094177, + 0.02597120590507984, + 0.02285279519855976, + 0.008649211376905441, + 0.010597443208098412, + 0.0089816739782691, + 0.008942087180912495, + 0.014407753944396973, + 0.007753143087029457, + 0.007326986640691757, + 0.007278898265212774, + 0.005277121439576149, + 0.005921253003180027, + 0.007050648797303438, + 0.011989052407443523, + 0.015626579523086548, + 0.011656330898404121, + 0.009736420586705208, + 0.011671473272144794, + 0.015425856225192547, + 0.013059011660516262, + 0.02160641737282276, + 0.005105027928948402, + 0.007619325537234545, + 0.006742799188941717, + 0.007969258353114128, + 0.006449888925999403, + 0.009713378734886646, + 0.014830127358436584, + 0.008858686313033104, + 0.009119043126702309, + 0.004419115372002125, + 0.011611109599471092, + 0.011234806850552559, + 0.006815243046730757, + 0.011934736743569374, + 0.009009488858282566, + 0.0110621964558959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.001522573409602046 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067332289181649685 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.0067332289181649685, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067332289181649685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1828659027814865, + 0.14215175807476044, + 0.0984034314751625, + 0.12561661005020142, + 0.10784303396940231, + 0.14448456466197968, + 0.05301240086555481, + 0.054299790412187576, + 0.045504845678806305, + 0.04581763595342636 + ], + "credit_map": [ + 0.03474735841155052, + 0.02338472753763199, + 0.02756618522107601, + 0.015456555411219597, + 0.01834503374993801, + 0.033347610384225845, + 0.013075966387987137, + 0.016942456364631653, + 0.011459968984127045, + 0.01536452304571867, + 0.015099595300853252, + 0.013641712255775928, + 0.03350518271327019, + 0.014233672060072422, + 0.02451697364449501, + 0.014330129139125347, + 0.007915815338492393, + 0.00896124355494976, + 0.01063562836498022, + 0.012179834768176079, + 0.012734035961329937, + 0.0172364953905344, + 0.013180805370211601, + 0.015559566207230091, + 0.0074173277243971825, + 0.008810334838926792, + 0.02301635406911373, + 0.02561851032078266, + 0.008941225707530975, + 0.014239737764000893, + 0.018618213012814522, + 0.018954899162054062, + 0.008082674816250801, + 0.00940696056932211, + 0.02630148082971573, + 0.014044294133782387, + 0.018965255469083786, + 0.006908718030899763, + 0.013702367432415485, + 0.010431280359625816, + 0.007558274082839489, + 0.019712883979082108, + 0.015431084670126438, + 0.043188225477933884, + 0.03152921050786972, + 0.007582248654216528, + 0.012114915996789932, + 0.007367722224444151, + 0.005825255531817675, + 0.013190275989472866, + 0.0065443930216133595, + 0.005629193503409624, + 0.00572890555486083, + 0.0050254277884960175, + 0.005564668215811253, + 0.00550428219139576, + 0.008024896495044231, + 0.008343423716723919, + 0.007520950399339199, + 0.005462048575282097, + 0.008285933174192905, + 0.005411557387560606, + 0.005431936122477055, + 0.005819041281938553, + 0.0032016239129006863, + 0.005037941038608551, + 0.005551902577280998, + 0.005334349349141121, + 0.0043436079286038876, + 0.005762890446931124, + 0.007653198204934597, + 0.008619333617389202, + 0.007441824302077293, + 0.004019256215542555, + 0.008032289333641529, + 0.0045486195012927055, + 0.004820771049708128, + 0.004376583732664585, + 0.004761054180562496, + 0.007817234843969345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0067332289181649685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic sequence imbued with dynamism and tension, composed of a series of rapid, violent edits, opens with a turbulent long shot of a weathered, dark wooden fishing boat bobbing violently on a turbulent, gray-black sea. Under a gloomy, stormy sky, waves crash against the boat like mountains. Then, through abrupt jump cuts, the scene switches back and forth between two striking close-ups: one is an extreme close-up of thorny, dark green vines, as thick as pythons, tightly wrapped around the wet side and mast, nearly strangling the boat. The vines, covered in sharp black thorns, are stretched straight under the immense tension, and even tiny splinters can be seen breaking from the squeezed planking. The other, captured from an extremely low angle, captures a massive, white-foaming wave crashing against the side of the boat, instantly exploding into a cloud of water droplets and mist, each droplet appearing crystal clear and cold in the dim light. The entire sequence is gloomy in color, dominated by deep blue, dark green, and the dark brown of wet wood, with only the white foam of the waves providing a harsh contrast. The rapid editing rhythm combined with the helpless tossing of the hull creates a tense atmosphere of struggle and despair, surrounded by the violence of nature.", + "seed": 22037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22037_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": -0.04051353125832975, + "vlm_logprob": -0.04051353125832975, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04051353125832975, + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11090794205665588 + ], + "opposite_logprob_tokens": [ + -2.360908031463623 + ], + "logprob_reward": -0.11090794205665588, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11090794205665588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11354966461658478, + 0.1679488718509674, + 0.11831849813461304, + 0.0918491929769516, + 0.08875937014818192, + 0.057553619146347046, + 0.2446104884147644, + 0.033990733325481415, + 0.024399546906352043, + 0.059020012617111206 + ], + "credit_map": [ + 0.014342759735882282, + 0.0165094006806612, + 0.00995517149567604, + 0.015744343400001526, + 0.013405367732048035, + 0.008478820323944092, + 0.022711370140314102, + 0.012402432970702648, + 0.009427931159734726, + 0.010365933179855347, + 0.015326122753322124, + 0.042652588337659836, + 0.00878799520432949, + 0.012465064413845539, + 0.041679173707962036, + 0.02724405564367771, + 0.005178418941795826, + 0.005044474266469479, + 0.005556653253734112, + 0.009169708006083965, + 0.004011841490864754, + 0.004427013453096151, + 0.04277455806732178, + 0.042155832052230835, + 0.008049373514950275, + 0.008307953365147114, + 0.017179643735289574, + 0.012654414400458336, + 0.012296930886805058, + 0.009517410770058632, + 0.013058873824775219, + 0.010784588754177094, + 0.007576320320367813, + 0.029583150520920753, + 0.004429585300385952, + 0.005277097690850496, + 0.026109887287020683, + 0.00707503454759717, + 0.0033442156855016947, + 0.005364078097045422, + 0.006245662923902273, + 0.007072914857417345, + 0.005804093554615974, + 0.005498243495821953, + 0.007553562521934509, + 0.010149321518838406, + 0.007898381911218166, + 0.007331442553550005, + 0.010519393719732761, + 0.10304684191942215, + 0.003531401976943016, + 0.005945501383394003, + 0.10257983207702637, + 0.010322281159460545, + 0.005488932598382235, + 0.0031763073056936264, + 0.003049601335078478, + 0.003769921138882637, + 0.0031354385428130627, + 0.002543862909078598, + 0.008405293338000774, + 0.008446370251476765, + 0.002837758045643568, + 0.001802490558475256, + 0.0031237155199050903, + 0.003773638978600502, + 0.0029670658987015486, + 0.0034385144244879484, + 0.0022487821988761425, + 0.00329115753993392, + 0.002736810827627778, + 0.002819861751049757, + 0.004819836001843214, + 0.006789028644561768, + 0.007004505023360252, + 0.007533346302807331, + 0.01023409515619278, + 0.008558753877878189, + 0.007935796864330769, + 0.006144648417830467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.11090794205665588 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011124728247523308 + ], + "opposite_logprob_tokens": [ + -4.511124610900879 + ], + "logprob_reward": -0.011124728247523308, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011124728247523308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05192326381802559, + 0.054215461015701294, + 0.11226467788219452, + 0.04901539534330368, + 0.07744016498327255, + 0.02849762886762619, + 0.4543779492378235, + 0.0742744505405426, + 0.04984813928604126, + 0.048142824321985245 + ], + "credit_map": [ + 0.005130314733833075, + 0.004319564439356327, + 0.004894966259598732, + 0.00802557822316885, + 0.006574184633791447, + 0.004272885620594025, + 0.01195452269166708, + 0.006751249544322491, + 0.005834341049194336, + 0.002843135967850685, + 0.005643561016768217, + 0.015063971281051636, + 0.0029972377233207226, + 0.003972199745476246, + 0.01280282810330391, + 0.005058193113654852, + 0.005334743298590183, + 0.007124115247279406, + 0.0065698460675776005, + 0.0139671815559268, + 0.0039573246613144875, + 0.004337838385254145, + 0.03547070547938347, + 0.03550294041633606, + 0.004042765591293573, + 0.004592592362314463, + 0.0062024290673434734, + 0.0059940405189991, + 0.006690658628940582, + 0.0054573784582316875, + 0.01027492806315422, + 0.005760604981333017, + 0.01077208761125803, + 0.013529034331440926, + 0.007006064057350159, + 0.009321183897554874, + 0.012307269498705864, + 0.013195215724408627, + 0.005142384208738804, + 0.006166932638734579, + 0.0032952288165688515, + 0.003243488958105445, + 0.0036822075489908457, + 0.0034069635439664125, + 0.003790944116190076, + 0.004387235268950462, + 0.0036761644296348095, + 0.0030153996776789427, + 0.019027676433324814, + 0.18380825221538544, + 0.00917341373860836, + 0.01401042565703392, + 0.18176321685314178, + 0.02550785429775715, + 0.01260576955974102, + 0.008481369353830814, + 0.005247917026281357, + 0.00918660219758749, + 0.004693920258432627, + 0.004107118118554354, + 0.022335723042488098, + 0.018924782052636147, + 0.00485523184761405, + 0.004923165310174227, + 0.006435540039092302, + 0.0065625873394310474, + 0.005950550083070993, + 0.0077760168351233006, + 0.00638645701110363, + 0.004285871516913176, + 0.007316789589822292, + 0.00513433525338769, + 0.003986038733273745, + 0.005862669087946415, + 0.0069809104315936565, + 0.005843861028552055, + 0.007185956928879023, + 0.006186675280332565, + 0.007254209369421005, + 0.004842507187277079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.011124728247523308 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008283585309982 + ], + "opposite_logprob_tokens": [ + -2.695082902908325 + ], + "logprob_reward": -0.07008283585309982, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008283585309982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0557919405400753, + 0.05893833562731743, + 0.13884444534778595, + 0.1281890571117401, + 0.07544345408678055, + 0.0400988832116127, + 0.373283714056015, + 0.05738421902060509, + 0.0223272442817688, + 0.049698684364557266 + ], + "credit_map": [ + 0.005181130953133106, + 0.006909256335347891, + 0.005071257706731558, + 0.008787797763943672, + 0.0066728186793625355, + 0.004489436745643616, + 0.009966308251023293, + 0.008713933639228344, + 0.006151621229946613, + 0.004832407925277948, + 0.008099132217466831, + 0.014867328107357025, + 0.004300921224057674, + 0.002635631710290909, + 0.012157526798546314, + 0.005893767345696688, + 0.006060339044779539, + 0.009797564707696438, + 0.007558752316981554, + 0.016351301223039627, + 0.005247754510492086, + 0.005097751505672932, + 0.044207166880369186, + 0.04452381655573845, + 0.011225448921322823, + 0.010478498414158821, + 0.01153403427451849, + 0.018216487020254135, + 0.017369573935866356, + 0.014843116514384747, + 0.02620377019047737, + 0.018318133428692818, + 0.00687975250184536, + 0.022737517952919006, + 0.0029947657603770494, + 0.004712618421763182, + 0.02645529806613922, + 0.005496110767126083, + 0.0024828461464494467, + 0.0036845391150563955, + 0.004265447612851858, + 0.004822006914764643, + 0.003611184423789382, + 0.003409407567232847, + 0.005299481097608805, + 0.007035993039608002, + 0.006083299871534109, + 0.005572066176682711, + 0.01602686569094658, + 0.15928180515766144, + 0.004819804802536964, + 0.008071967400610447, + 0.1581558883190155, + 0.014467808417975903, + 0.008146953769028187, + 0.004312611185014248, + 0.003933959174901247, + 0.007490119896829128, + 0.0033169034868478775, + 0.0035787872038781643, + 0.016673430800437927, + 0.016734011471271515, + 0.003155594225972891, + 0.0025014118291437626, + 0.0030710676219314337, + 0.003029055427759886, + 0.0026627627667039633, + 0.003718065097928047, + 0.0026906048879027367, + 0.002425881801173091, + 0.002587220398709178, + 0.0021425869781523943, + 0.003228038316592574, + 0.004772624466568232, + 0.011022550985217094, + 0.009331835433840752, + 0.00634666858240962, + 0.006145766470581293, + 0.0043195392936468124, + 0.004531660117208958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.07008283585309982 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028278145473450422 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028278145473450422, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028278145473450422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11542344093322754, + 0.13428497314453125, + 0.11523520201444626, + 0.0679624080657959, + 0.11853443831205368, + 0.048145197331905365, + 0.2266235649585724, + 0.05672788619995117, + 0.05428454279899597, + 0.06277832388877869 + ], + "credit_map": [ + 0.015612802468240261, + 0.01428100187331438, + 0.011252863332629204, + 0.017688700929284096, + 0.011387440375983715, + 0.00938404444605112, + 0.022708354517817497, + 0.013108239509165287, + 0.009043969213962555, + 0.009157749824225903, + 0.012968004681169987, + 0.03581506386399269, + 0.006659095175564289, + 0.009536582045257092, + 0.03183620795607567, + 0.01926829107105732, + 0.004785181023180485, + 0.005342674907296896, + 0.005526663735508919, + 0.01008250005543232, + 0.00414319010451436, + 0.0045600770972669125, + 0.040560364723205566, + 0.04023453965783119, + 0.006243272218853235, + 0.006865252275019884, + 0.012877010740339756, + 0.0068865022622048855, + 0.009760096669197083, + 0.007234017364680767, + 0.01002174150198698, + 0.008074520155787468, + 0.007977545261383057, + 0.045829080045223236, + 0.0037339215632528067, + 0.0037679672241210938, + 0.043457936495542526, + 0.007912719622254372, + 0.0024397755041718483, + 0.0034154891036450863, + 0.004837095271795988, + 0.0065779248252511024, + 0.00560745457187295, + 0.006757059600204229, + 0.006937459111213684, + 0.00627897260710597, + 0.005748558323830366, + 0.0054006753489375114, + 0.012294499203562737, + 0.07767700403928757, + 0.006333524361252785, + 0.014239580370485783, + 0.07919985055923462, + 0.022712720558047295, + 0.008809356018900871, + 0.005357029382139444, + 0.0033162571489810944, + 0.005462264642119408, + 0.004475864581763744, + 0.0034213238395750523, + 0.016275662928819656, + 0.01721806451678276, + 0.0028962865471839905, + 0.003662157105281949, + 0.006918550934642553, + 0.009160579182207584, + 0.006172614172101021, + 0.007089931517839432, + 0.004523879382759333, + 0.00737785967066884, + 0.0074204313568770885, + 0.00562069658190012, + 0.0053081694059073925, + 0.007639378309249878, + 0.008444536477327347, + 0.008370437659323215, + 0.009848810732364655, + 0.008674316108226776, + 0.008248942904174328, + 0.006243731826543808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.0028278145473450422 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624335587024689 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624335587024689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624335587024689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04516647011041641, + 0.07864760607481003, + 0.05745667219161987, + 0.023638224229216576, + 0.08953798562288284, + 0.03109634853899479, + 0.563018798828125, + 0.059628166258335114, + 0.032640427350997925, + 0.019169310107827187 + ], + "credit_map": [ + 0.004832676146179438, + 0.004328218754380941, + 0.004919502884149551, + 0.007619783747941256, + 0.0041604978032410145, + 0.0033638959284871817, + 0.010831245221197605, + 0.005110651254653931, + 0.009070019237697124, + 0.005442900583148003, + 0.009422458708286285, + 0.02388642355799675, + 0.004362435545772314, + 0.0036145891062915325, + 0.017985621467232704, + 0.004863155540078878, + 0.0032713674008846283, + 0.004973685834556818, + 0.00478277588263154, + 0.00798027589917183, + 0.0021927289199084044, + 0.002930370857939124, + 0.016068508848547935, + 0.015256958082318306, + 0.002479993272572756, + 0.0038139501120895147, + 0.004072750452905893, + 0.0026359728071838617, + 0.002491000574082136, + 0.002296286867931485, + 0.003293678630143404, + 0.002554589184001088, + 0.011233207769691944, + 0.020917557179927826, + 0.006564669776707888, + 0.0074035110883414745, + 0.029168028384447098, + 0.0071006519719958305, + 0.0030561492312699556, + 0.004094218835234642, + 0.003492556745186448, + 0.003909421619027853, + 0.0033785798586905003, + 0.0034829420037567616, + 0.004173638299107552, + 0.0052037970162928104, + 0.0040233745239675045, + 0.0034320391714572906, + 0.023446423932909966, + 0.23385091125965118, + 0.009255492128431797, + 0.014686232432723045, + 0.231558158993721, + 0.02798932045698166, + 0.013432363979518414, + 0.00879989005625248, + 0.0058015794493258, + 0.011144991964101791, + 0.004670616239309311, + 0.004027923569083214, + 0.014389224350452423, + 0.011592861264944077, + 0.0045170472003519535, + 0.0034839275758713484, + 0.0037622360978275537, + 0.0050054555758833885, + 0.003945525269955397, + 0.003465944202616811, + 0.003786243498325348, + 0.004653480369597673, + 0.0043051284737885, + 0.0037164120003581047, + 0.001564886886626482, + 0.0021610683761537075, + 0.0020215888507664204, + 0.00281404796987772, + 0.002837137086316943, + 0.0028358122799545527, + 0.0029212087392807007, + 0.0020135585218667984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.007624335587024689 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic sequence imbued with dynamism and tension, composed of a series of rapid, violent edits, opens with a turbulent long shot of a weathered, dark wooden fishing boat bobbing violently on a turbulent, gray-black sea. Under a gloomy, stormy sky, waves crash against the boat like mountains. Then, through abrupt jump cuts, the scene switches back and forth between two striking close-ups: one is an extreme close-up of thorny, dark green vines, as thick as pythons, tightly wrapped around the wet side and mast, nearly strangling the boat. The vines, covered in sharp black thorns, are stretched straight under the immense tension, and even tiny splinters can be seen breaking from the squeezed planking. The other, captured from an extremely low angle, captures a massive, white-foaming wave crashing against the side of the boat, instantly exploding into a cloud of water droplets and mist, each droplet appearing crystal clear and cold in the dim light. The entire sequence is gloomy in color, dominated by deep blue, dark green, and the dark brown of wet wood, with only the white foam of the waves providing a harsh contrast. The rapid editing rhythm combined with the helpless tossing of the hull creates a tense atmosphere of struggle and despair, surrounded by the violence of nature.", + "seed": 22038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22038_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": -0.03039502096362412, + "vlm_logprob": -0.03039502096362412, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.03039502096362412, + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06606034934520721 + ], + "opposite_logprob_tokens": [ + -3.8160603046417236 + ], + "logprob_reward": -0.06606034934520721, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.06606034934520721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24778246879577637, + 0.13536590337753296, + 0.07819288223981857, + 0.08443142473697662, + 0.14777924120426178, + 0.03372616693377495, + 0.0525246188044548, + 0.0876203253865242, + 0.09705322235822678, + 0.0355236679315567 + ], + "credit_map": [ + 0.03438480943441391, + 0.02602500468492508, + 0.028604678809642792, + 0.023802297189831734, + 0.03720773383975029, + 0.026798389852046967, + 0.022753799334168434, + 0.04820576682686806, + 0.012803505174815655, + 0.013007045723497868, + 0.014997337013483047, + 0.010026141069829464, + 0.03480130806565285, + 0.0328301265835762, + 0.009107540361583233, + 0.007792921271175146, + 0.007137649692595005, + 0.006506969686597586, + 0.00855328980833292, + 0.007383949123322964, + 0.016503077000379562, + 0.01596350036561489, + 0.00808017235249281, + 0.00806428026407957, + 0.007696531713008881, + 0.006486125290393829, + 0.011805234476923943, + 0.011393397115170956, + 0.008861804381012917, + 0.00547640398144722, + 0.0182002242654562, + 0.014511710964143276, + 0.010509146377444267, + 0.015246025286614895, + 0.024841032922267914, + 0.031266096979379654, + 0.00869880523532629, + 0.00948502216488123, + 0.024454547092318535, + 0.02327858656644821, + 0.0026192667428404093, + 0.0032010048162192106, + 0.005342039745301008, + 0.002624610671773553, + 0.005968708544969559, + 0.0042121754959225655, + 0.004695347975939512, + 0.005063018295913935, + 0.0043309153988957405, + 0.009752324782311916, + 0.004766115453094244, + 0.005558363627642393, + 0.006418434903025627, + 0.006481176242232323, + 0.008598385378718376, + 0.006618908140808344, + 0.015450788661837578, + 0.017642224207520485, + 0.0124236224219203, + 0.007339330855756998, + 0.007425074465572834, + 0.00842670351266861, + 0.009327925741672516, + 0.00958466250449419, + 0.015475932508707047, + 0.01315449457615614, + 0.009353871457278728, + 0.00765613978728652, + 0.020202217623591423, + 0.01656734198331833, + 0.006337495055049658, + 0.008305735886096954, + 0.004240879788994789, + 0.004689258988946676, + 0.004654575139284134, + 0.0028100842610001564, + 0.003994101192802191, + 0.0068048895336687565, + 0.004403258208185434, + 0.003926623612642288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06606034934520721 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012582560069859028 + ], + "opposite_logprob_tokens": [ + -4.387582778930664 + ], + "logprob_reward": -0.012582560069859028, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012582560069859028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07133453339338303, + 0.16590824723243713, + 0.091937355697155, + 0.07307703047990799, + 0.08638621121644974, + 0.14649628102779388, + 0.10589152574539185, + 0.05450305715203285, + 0.08892375975847244, + 0.11554199457168579 + ], + "credit_map": [ + 0.00705087510868907, + 0.008482436649501324, + 0.006267643999308348, + 0.0074621341191232204, + 0.017316844314336777, + 0.010123608633875847, + 0.006854657083749771, + 0.007776338141411543, + 0.008942951448261738, + 0.011674794368445873, + 0.007734110578894615, + 0.007001217920333147, + 0.059386834502220154, + 0.058211568742990494, + 0.0068021053448319435, + 0.006154660135507584, + 0.00889922771602869, + 0.006279307883232832, + 0.010018394328653812, + 0.006514727137982845, + 0.01771709695458412, + 0.01797446608543396, + 0.009935073554515839, + 0.014599055051803589, + 0.00688053946942091, + 0.006851102691143751, + 0.009656022302806377, + 0.011807584203779697, + 0.005651333834975958, + 0.005696854088455439, + 0.014066539704799652, + 0.012467047199606895, + 0.005751532502472401, + 0.009948273189365864, + 0.012836695648729801, + 0.012559641152620316, + 0.005555249750614166, + 0.00693528400734067, + 0.015874842181801796, + 0.016924694180488586, + 0.008905675262212753, + 0.015129386447370052, + 0.023414162918925285, + 0.012245354242622852, + 0.019804583862423897, + 0.022224975749850273, + 0.02256857417523861, + 0.022203579545021057, + 0.01514432393014431, + 0.01598910242319107, + 0.008754773065447807, + 0.011670074425637722, + 0.01562261302024126, + 0.018017353489995003, + 0.01082672830671072, + 0.009866561740636826, + 0.006401284132152796, + 0.00965811312198639, + 0.007918250747025013, + 0.004624028690159321, + 0.0059171998873353004, + 0.005445458926260471, + 0.005813063122332096, + 0.008725658990442753, + 0.014755060896277428, + 0.011804758571088314, + 0.007889497093856335, + 0.007049059495329857, + 0.015976669266819954, + 0.011369188316166401, + 0.010463515296578407, + 0.009616012685000896, + 0.014205188490450382, + 0.021118558943271637, + 0.010200967080891132, + 0.011677554808557034, + 0.014274878427386284, + 0.020293116569519043, + 0.01269675325602293, + 0.011074986308813095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.012582560069859028 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207999214529991 + ], + "opposite_logprob_tokens": [ + -2.812079906463623 + ], + "logprob_reward": -0.06207999214529991, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207999214529991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20383402705192566, + 0.1750461310148239, + 0.06849011778831482, + 0.08768629282712936, + 0.17257335782051086, + 0.07022252678871155, + 0.06396640092134476, + 0.05499543249607086, + 0.02644171565771103, + 0.07674398273229599 + ], + "credit_map": [ + 0.029916035011410713, + 0.021630816161632538, + 0.019324569031596184, + 0.015573662705719471, + 0.03421139717102051, + 0.028618115931749344, + 0.0187035221606493, + 0.03585593029856682, + 0.007638337556272745, + 0.005521118640899658, + 0.0067312028259038925, + 0.005700243636965752, + 0.07067348062992096, + 0.06815066188573837, + 0.005451603792607784, + 0.00517948716878891, + 0.006708356551826, + 0.0038879583589732647, + 0.008036158978939056, + 0.012135105207562447, + 0.009016680531203747, + 0.008632725104689598, + 0.009668507613241673, + 0.010404623113572598, + 0.004328023176640272, + 0.009338189847767353, + 0.014823857694864273, + 0.022578272968530655, + 0.004994513466954231, + 0.0058207400143146515, + 0.012740951031446457, + 0.013061749748885632, + 0.009263124316930771, + 0.017514575272798538, + 0.026966644451022148, + 0.033717457205057144, + 0.010090939700603485, + 0.011191736906766891, + 0.032387398183345795, + 0.03144149109721184, + 0.0033135784324258566, + 0.007735001388937235, + 0.009398827329277992, + 0.0083863390609622, + 0.005749704781919718, + 0.00736432196572423, + 0.016801664605736732, + 0.011473095044493675, + 0.01237895991653204, + 0.008057028986513615, + 0.003883058438077569, + 0.004134676419198513, + 0.01006963849067688, + 0.015470336191356182, + 0.005551541689783335, + 0.004421166144311428, + 0.00861078966408968, + 0.012126636691391468, + 0.006443866528570652, + 0.0047624665312469006, + 0.006457965355366468, + 0.0062650335021317005, + 0.004501557908952236, + 0.005827120039612055, + 0.004044061992317438, + 0.003978407941758633, + 0.0024129534140229225, + 0.0025867358781397343, + 0.0039437212981283665, + 0.0041140117682516575, + 0.0028779220301657915, + 0.002483904827386141, + 0.010271328501403332, + 0.014668077230453491, + 0.006458447314798832, + 0.006439140997827053, + 0.008476783521473408, + 0.015270303003489971, + 0.008412007242441177, + 0.0067479051649570465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.06207999214529991 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036286956164985895 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.0036286956164985895, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036286956164985895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759835034608841, + 0.1970592737197876, + 0.04201672971248627, + 0.05022447928786278, + 0.12931768596172333, + 0.05667132884263992, + 0.09204430878162384, + 0.05619443580508232, + 0.17429257929325104, + 0.02619572915136814 + ], + "credit_map": [ + 0.020314613357186317, + 0.01664670556783676, + 0.033236823976039886, + 0.023090841248631477, + 0.024034086614847183, + 0.02263835445046425, + 0.011931706219911575, + 0.024090372025966644, + 0.015467983670532703, + 0.014425667934119701, + 0.017423709854483604, + 0.012134699150919914, + 0.06087532639503479, + 0.05882962420582771, + 0.009814916178584099, + 0.008087347261607647, + 0.003932665102183819, + 0.0036889559123665094, + 0.00514558982104063, + 0.006106918212026358, + 0.005856164265424013, + 0.005266860593110323, + 0.004934299737215042, + 0.007085275370627642, + 0.004300812259316444, + 0.004949838854372501, + 0.007481037639081478, + 0.01091956440359354, + 0.0034780444111675024, + 0.004498640540987253, + 0.007132259197533131, + 0.007464281283318996, + 0.010129706934094429, + 0.016633328050374985, + 0.020170586183667183, + 0.025506986305117607, + 0.007810742594301701, + 0.00883450172841549, + 0.016991565003991127, + 0.023240262642502785, + 0.004771315027028322, + 0.004976232536137104, + 0.010367631912231445, + 0.0059721521101891994, + 0.006908641196787357, + 0.007464724592864513, + 0.0068361349403858185, + 0.009374494664371014, + 0.015308554284274578, + 0.009641535580158234, + 0.005777096375823021, + 0.005396505817770958, + 0.021054673939943314, + 0.0211727824062109, + 0.008217751048505306, + 0.0054754032753407955, + 0.006261180154979229, + 0.010587194934487343, + 0.010362157598137856, + 0.005068462807685137, + 0.004970885347574949, + 0.005659535061568022, + 0.007079713512212038, + 0.006205309182405472, + 0.0437791645526886, + 0.03436335176229477, + 0.012177086435258389, + 0.011410954408347607, + 0.019848471507430077, + 0.02772384323179722, + 0.011548230424523354, + 0.013441461138427258, + 0.0036024092696607113, + 0.004465300124138594, + 0.002102653495967388, + 0.0016383141046389937, + 0.0027059605345129967, + 0.0053827413357794285, + 0.0033650833647698164, + 0.0029332663398236036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0036286956164985895 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076235076412558556 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.0076235076412558556, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076235076412558556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15931200981140137, + 0.15869922935962677, + 0.09268195927143097, + 0.07956334948539734, + 0.11326466500759125, + 0.0703638568520546, + 0.12352597713470459, + 0.10059834271669388, + 0.06663288921117783, + 0.035357724875211716 + ], + "credit_map": [ + 0.021915646269917488, + 0.0173668060451746, + 0.014794372022151947, + 0.014203007332980633, + 0.032418567687273026, + 0.022508397698402405, + 0.012041820213198662, + 0.024063391610980034, + 0.008580047637224197, + 0.014246425591409206, + 0.009280279278755188, + 0.008551465347409248, + 0.052754539996385574, + 0.050542183220386505, + 0.007838299497961998, + 0.006905984599143267, + 0.011137522757053375, + 0.008296841755509377, + 0.010350693017244339, + 0.008884579874575138, + 0.01740078441798687, + 0.016486624255776405, + 0.009541728533804417, + 0.010583185590803623, + 0.00446140905842185, + 0.0060993945226073265, + 0.012158943340182304, + 0.018134446814656258, + 0.007333011832088232, + 0.008872607722878456, + 0.012339487671852112, + 0.010164042003452778, + 0.0069480035454034805, + 0.009637990966439247, + 0.022589217871427536, + 0.0212943684309721, + 0.006917162332683802, + 0.009075826965272427, + 0.01562771201133728, + 0.02117437683045864, + 0.004490962252020836, + 0.0077170273289084435, + 0.00867676455527544, + 0.005017691291868687, + 0.009112351574003696, + 0.007889836095273495, + 0.014750638045370579, + 0.012708589434623718, + 0.017337193712592125, + 0.0173297431319952, + 0.01065092347562313, + 0.011694321408867836, + 0.023180820047855377, + 0.020957335829734802, + 0.011690995655953884, + 0.010684638284146786, + 0.01017939206212759, + 0.01425893697887659, + 0.018902979791164398, + 0.007138041779398918, + 0.010491461493074894, + 0.009915742091834545, + 0.015608702786266804, + 0.014103089459240437, + 0.012277533300220966, + 0.012746678665280342, + 0.005458377301692963, + 0.005599956959486008, + 0.009766656905412674, + 0.007844097912311554, + 0.007195389363914728, + 0.0057441978715360165, + 0.0038733312394469976, + 0.004777464084327221, + 0.0043977005407214165, + 0.004033926408737898, + 0.0037782606668770313, + 0.005611964967101812, + 0.004750297404825687, + 0.00413477560505271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0076235076412558556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic sequence imbued with dynamism and tension, composed of a series of rapid, violent edits, opens with a turbulent long shot of a weathered, dark wooden fishing boat bobbing violently on a turbulent, gray-black sea. Under a gloomy, stormy sky, waves crash against the boat like mountains. Then, through abrupt jump cuts, the scene switches back and forth between two striking close-ups: one is an extreme close-up of thorny, dark green vines, as thick as pythons, tightly wrapped around the wet side and mast, nearly strangling the boat. The vines, covered in sharp black thorns, are stretched straight under the immense tension, and even tiny splinters can be seen breaking from the squeezed planking. The other, captured from an extremely low angle, captures a massive, white-foaming wave crashing against the side of the boat, instantly exploding into a cloud of water droplets and mist, each droplet appearing crystal clear and cold in the dim light. The entire sequence is gloomy in color, dominated by deep blue, dark green, and the dark brown of wet wood, with only the white foam of the waves providing a harsh contrast. The rapid editing rhythm combined with the helpless tossing of the hull creates a tense atmosphere of struggle and despair, surrounded by the violence of nature.", + "seed": 22039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_2/index_7006_seed_22039_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": -0.030744917038828135, + "vlm_logprob": -0.030744917038828135, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.030744917038828135, + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06456495076417923 + ], + "opposite_logprob_tokens": [ + -3.0645649433135986 + ], + "logprob_reward": -0.06456495076417923, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06456495076417923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21816888451576233, + 0.02795456536114216, + 0.08548328280448914, + 0.057802218943834305, + 0.1395532339811325, + 0.09674953669309616, + 0.02784295566380024, + 0.16608114540576935, + 0.05860159918665886, + 0.12176257371902466 + ], + "credit_map": [ + 0.030766652897000313, + 0.026355881243944168, + 0.021386723965406418, + 0.03109639510512352, + 0.024035438895225525, + 0.022655608132481575, + 0.03450878709554672, + 0.02736339345574379, + 0.003329632105305791, + 0.004214255139231682, + 0.0038777778390794992, + 0.0033013878855854273, + 0.004434940870851278, + 0.003102991497144103, + 0.0029898532666265965, + 0.0027037253603339195, + 0.00780866714194417, + 0.014827496372163296, + 0.012682334519922733, + 0.007812206633388996, + 0.01748535968363285, + 0.01061750203371048, + 0.008756276220083237, + 0.00549344252794981, + 0.007461334113031626, + 0.00877821259200573, + 0.007228956092149019, + 0.00916732382029295, + 0.0077888453379273415, + 0.005752203520387411, + 0.0057008774019777775, + 0.005924469791352749, + 0.02220403403043747, + 0.03600187227129936, + 0.009639644995331764, + 0.012024737894535065, + 0.02033478207886219, + 0.023122642189264297, + 0.00912387203425169, + 0.007101642433553934, + 0.012383680790662766, + 0.012014077976346016, + 0.009695137850940228, + 0.012628833763301373, + 0.019061371684074402, + 0.008784964680671692, + 0.010474256239831448, + 0.011707212775945663, + 0.001984864007681608, + 0.0028966323006898165, + 0.0032007808331400156, + 0.004779423121362925, + 0.004063824191689491, + 0.004887108691036701, + 0.0029023014940321445, + 0.003128021489828825, + 0.017438096925616264, + 0.011464637704193592, + 0.0064122541807591915, + 0.025468554347753525, + 0.013799077831208706, + 0.019125070422887802, + 0.05428726226091385, + 0.018086178228259087, + 0.009988452307879925, + 0.009953556582331657, + 0.006150657776743174, + 0.00665651960298419, + 0.006405272986739874, + 0.005865668877959251, + 0.005280388053506613, + 0.008301081135869026, + 0.013843739405274391, + 0.027726715430617332, + 0.014886585995554924, + 0.015831565484404564, + 0.012378353625535965, + 0.011891305446624756, + 0.01331515982747078, + 0.011889149434864521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.06456495076417923 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023338036611676216 + ], + "opposite_logprob_tokens": [ + -3.7733380794525146 + ], + "logprob_reward": -0.023338036611676216, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023338036611676216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04362217336893082, + 0.05985187366604805, + 0.04977541044354439, + 0.07237806916236877, + 0.24990174174308777, + 0.09572417289018631, + 0.09117116034030914, + 0.1502165049314499, + 0.039017435163259506, + 0.14834146201610565 + ], + "credit_map": [ + 0.0059334710240364075, + 0.00536468718200922, + 0.003869340056553483, + 0.00622022757306695, + 0.005222483072429895, + 0.005507999565452337, + 0.005980898626148701, + 0.005523061379790306, + 0.007908429950475693, + 0.00846483837813139, + 0.004850327502936125, + 0.008843700401484966, + 0.011292366310954094, + 0.005343652796000242, + 0.006907081697136164, + 0.006241476628929377, + 0.005545501597225666, + 0.007865420542657375, + 0.007544463500380516, + 0.005229257047176361, + 0.007566177286207676, + 0.008598792366683483, + 0.0038577208761125803, + 0.0035680721048265696, + 0.007271529175341129, + 0.008057261817157269, + 0.0060142334550619125, + 0.009060312062501907, + 0.013591443188488483, + 0.011870644986629486, + 0.008598096668720245, + 0.007914537563920021, + 0.052874233573675156, + 0.056998688727617264, + 0.017417175695300102, + 0.021607492119073868, + 0.036571208387613297, + 0.03706919401884079, + 0.013788970187306404, + 0.01357476506382227, + 0.011246057227253914, + 0.013668455183506012, + 0.009352568536996841, + 0.00966708641499281, + 0.020745953544974327, + 0.0077896821312606335, + 0.011173027567565441, + 0.012081328779459, + 0.0075178141705691814, + 0.012947069481015205, + 0.008792780339717865, + 0.012750988826155663, + 0.010326393879950047, + 0.015817414969205856, + 0.00978333130478859, + 0.013235361315310001, + 0.015284471213817596, + 0.022414740175008774, + 0.01142448652535677, + 0.013954957015812397, + 0.02148469164967537, + 0.030052008107304573, + 0.01641760766506195, + 0.019183533266186714, + 0.003754038130864501, + 0.004600043408572674, + 0.002883274806663394, + 0.007028782274574041, + 0.00460459291934967, + 0.0037374203093349934, + 0.00596044072881341, + 0.006448838859796524, + 0.018840961158275604, + 0.041534174233675, + 0.020373467355966568, + 0.016376592218875885, + 0.012483753263950348, + 0.013026867993175983, + 0.01322541106492281, + 0.012480227276682854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.023338036611676216 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054996609687805176 + ], + "opposite_logprob_tokens": [ + -2.9299964904785156 + ], + "logprob_reward": -0.054996609687805176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054996609687805176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10502473264932632, + 0.041912440210580826, + 0.07141226530075073, + 0.06311438977718353, + 0.15849027037620544, + 0.10251635313034058, + 0.16256797313690186, + 0.16156253218650818, + 0.0553690642118454, + 0.07802996784448624 + ], + "credit_map": [ + 0.014703389257192612, + 0.012328182347118855, + 0.008117301389575005, + 0.013308633118867874, + 0.012275135144591331, + 0.018567660823464394, + 0.011771987192332745, + 0.013952442444860935, + 0.005795112811028957, + 0.00805519986897707, + 0.0025774547830224037, + 0.007349109742790461, + 0.006000848952680826, + 0.004451266489923, + 0.004459184594452381, + 0.0032242643646895885, + 0.00535923196002841, + 0.010778172872960567, + 0.00804996956139803, + 0.00663296552374959, + 0.011498902924358845, + 0.012349567376077175, + 0.011818600818514824, + 0.004924854263663292, + 0.007639605086296797, + 0.008284928277134895, + 0.010953737422823906, + 0.008335617370903492, + 0.008539513684809208, + 0.006135343573987484, + 0.008125090040266514, + 0.005100554786622524, + 0.024753687903285027, + 0.04224833846092224, + 0.009443606249988079, + 0.012777340598404408, + 0.023477518931031227, + 0.02675156109035015, + 0.009402690455317497, + 0.009635518305003643, + 0.01289751660078764, + 0.013775400817394257, + 0.011638223193585873, + 0.013380855321884155, + 0.019571606069803238, + 0.009216832928359509, + 0.010357880964875221, + 0.011678031645715237, + 0.008814102970063686, + 0.007812112104147673, + 0.03566475957632065, + 0.06135157123208046, + 0.005235436372458935, + 0.00814498495310545, + 0.014931222423911095, + 0.020613783970475197, + 0.016574153676629066, + 0.023619545623660088, + 0.013651101849973202, + 0.019499000161886215, + 0.018035583198070526, + 0.02338830567896366, + 0.023609846830368042, + 0.023184997960925102, + 0.007998568005859852, + 0.007041092496365309, + 0.004906207788735628, + 0.006507665850222111, + 0.006015316117554903, + 0.005089465528726578, + 0.006677348632365465, + 0.011133400723338127, + 0.007709843572229147, + 0.011642133817076683, + 0.007432635873556137, + 0.013980651274323463, + 0.009136633016169071, + 0.007116585038602352, + 0.011431084014475346, + 0.009580397978425026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.054996609687805176 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032014804892241955 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.0032014804892241955, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032014804892241955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07261617481708527, + 0.031680695712566376, + 0.022607043385505676, + 0.04598207399249077, + 0.3693912923336029, + 0.20561742782592773, + 0.03314787521958351, + 0.11266163736581802, + 0.0421353280544281, + 0.06416044384241104 + ], + "credit_map": [ + 0.013091878034174442, + 0.0088536087423563, + 0.00801287591457367, + 0.010767418891191483, + 0.007382569834589958, + 0.0091773746535182, + 0.008327714167535305, + 0.00700274296104908, + 0.004319212399423122, + 0.003962712362408638, + 0.0026918496005237103, + 0.004312159959226847, + 0.0042554219253361225, + 0.004890428390353918, + 0.004037119913846254, + 0.0032117946539074183, + 0.00264148972928524, + 0.0029631920624524355, + 0.0024250932037830353, + 0.0018222390208393335, + 0.002408432774245739, + 0.003903209464624524, + 0.0050631435588002205, + 0.0013802462490275502, + 0.004145624116063118, + 0.0056774006225168705, + 0.0034238649532198906, + 0.007278688717633486, + 0.00797831080853939, + 0.006701617036014795, + 0.0055497800931334496, + 0.005226790904998779, + 0.06415563076734543, + 0.09519794583320618, + 0.024144181981682777, + 0.03122938983142376, + 0.05469191446900368, + 0.06389248371124268, + 0.01854642853140831, + 0.01753336377441883, + 0.027732636779546738, + 0.023263541981577873, + 0.02363619953393936, + 0.017986733466386795, + 0.039628300815820694, + 0.019010618329048157, + 0.02552250400185585, + 0.028836911544203758, + 0.002422620076686144, + 0.002241155132651329, + 0.0062914034351706505, + 0.010395454242825508, + 0.0027425207663327456, + 0.001978629967197776, + 0.004365067929029465, + 0.0027110285591334105, + 0.011988624930381775, + 0.009787023067474365, + 0.005409378558397293, + 0.017580457031726837, + 0.007833806797862053, + 0.010921239852905273, + 0.03626180440187454, + 0.01287932600826025, + 0.00833797361701727, + 0.005883936304599047, + 0.004402768332511187, + 0.004341760650277138, + 0.004780405666679144, + 0.0044150035828351974, + 0.0047430358827114105, + 0.005230449140071869, + 0.007569917477667332, + 0.015121512115001678, + 0.008615560829639435, + 0.00849279947578907, + 0.005430786870419979, + 0.0061793336644768715, + 0.006894299760460854, + 0.005856246221810579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.0032014804892241955 + }, + { + "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 turbulent long shot of a weathered, dark wooden fishing boat on a gray-black sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there abrupt jump cuts switching between an extreme close-up of thorny dark green vines and a low-angle shot of a massive white-foaming wave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the vines shown as thick as pythons, wrapped around the boat's side and mast, with visible splinters breaking from the planking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the color palette dominated by deep blue, dark green, and dark brown wet wood, with only white foam providing contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video composed of a series of rapid, violent edits creating a tense atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076235076412558556 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.0076235076412558556, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076235076412558556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05664785951375961, + 0.04724859073758125, + 0.09863497316837311, + 0.030762100592255592, + 0.10471680015325546, + 0.1122891753911972, + 0.026577183976769447, + 0.4411512613296509, + 0.024259071797132492, + 0.057712964713573456 + ], + "credit_map": [ + 0.009286090731620789, + 0.007751446217298508, + 0.004849926568567753, + 0.006616737227886915, + 0.006748759187757969, + 0.006474209949374199, + 0.008676035329699516, + 0.006244652904570103, + 0.006264645606279373, + 0.005894845817238092, + 0.003369547426700592, + 0.006364381872117519, + 0.008875406347215176, + 0.007046030834317207, + 0.005186247173696756, + 0.004247484728693962, + 0.00779114942997694, + 0.009746472351253033, + 0.009218699298799038, + 0.006257737521082163, + 0.009187141433358192, + 0.024031037464737892, + 0.02516121044754982, + 0.007241531275212765, + 0.003286690218374133, + 0.004034578800201416, + 0.0037568106781691313, + 0.005716816056519747, + 0.004126762505620718, + 0.0035984304267913103, + 0.0026329797692596912, + 0.003609033767133951, + 0.02175690419971943, + 0.02102784626185894, + 0.008340663276612759, + 0.009369894862174988, + 0.013583658263087273, + 0.013547456823289394, + 0.00969828199595213, + 0.007392102852463722, + 0.013813154771924019, + 0.01320947241038084, + 0.012588315643370152, + 0.01562272198498249, + 0.022499341517686844, + 0.008667782880365849, + 0.013192924670875072, + 0.012695461511611938, + 0.0015224713133648038, + 0.002048177644610405, + 0.0030863939318805933, + 0.004784559365361929, + 0.0017343645449727774, + 0.003809212241321802, + 0.002733173780143261, + 0.00685883080586791, + 0.04912581667304039, + 0.028795145452022552, + 0.01712772622704506, + 0.07349415868520737, + 0.026629064232110977, + 0.0349014475941658, + 0.16033749282360077, + 0.05074040964245796, + 0.0037359576672315598, + 0.0034322224091738462, + 0.0024259292986243963, + 0.002661831444129348, + 0.0022574798204004765, + 0.0024251393042504787, + 0.0027075251564383507, + 0.00461298692971468, + 0.00642599118873477, + 0.011764777824282646, + 0.0055994801223278046, + 0.00820854026824236, + 0.0060689011588692665, + 0.005801738705486059, + 0.0072626881301403046, + 0.00658085010945797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.0076235076412558556 + } + ], + "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/0000017/group_3/index_tensor([9680, 9680])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_tensor([9680, 9680])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b86235ac35788790926424cac4c5856f6311824f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_tensor([9680, 9680])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static wide-angle shot, rendered with a clean digital aesthetic, captures a poignant nighttime scene. Yuzuru Nishimiya, a young girl with short, light brown hair gently framing her face, stands in the center of a simple concrete pedestrian bridge. She wears a classic Japanese school uniform, consisting of a white, sailor-style shirt and a dark, pleated skirt, the fabric appearing almost black in the low light. The bright, cool glow of a large, full moon, just rising above the horizon, casts her figure into a stark, defined silhouette, while depicting delicate silver rims of light along her hair and shoulders. Below the bridge, a slender reflection of the moon shimmers on the deep, dark waters of a slow-flowing river. Her pose is one of profound stillness, her hands resting lightly on the cool metal railing, her gaze downward. Her facial expression is lost in the shadows, yet her body conveys a deep sense of quiet contemplation. The cool evening air is noticeably crisp, almost tangible, and in the distance the soft glow of city lights dissolves into a beautiful, softly warm-toned bokeh, creating a serene and melancholic atmosphere that isolates her in this introspective moment.", + "seed": 8801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8801_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": -0.010045753768645227, + "vlm_logprob": -0.010045753768645227, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.010045753768645227, + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.22659553587436676, + 0.0918385460972786, + 0.08468077331781387, + 0.07616196572780609, + 0.08681084215641022, + 0.03803621232509613, + 0.14251360297203064, + 0.08848151564598083, + 0.09616067260503769, + 0.06872037798166275 + ], + "credit_map": [ + 0.02812844328582287, + 0.02683459408581257, + 0.03395001217722893, + 0.017192712053656578, + 0.03302057087421417, + 0.030554894357919693, + 0.031530894339084625, + 0.02538338117301464, + 0.010156823322176933, + 0.007357897236943245, + 0.008401704020798206, + 0.004465687088668346, + 0.011641088873147964, + 0.014715536497533321, + 0.009353178553283215, + 0.02574661746621132, + 0.0068692839704453945, + 0.006692910101264715, + 0.011166314594447613, + 0.006322362460196018, + 0.010480174794793129, + 0.012413565069437027, + 0.009501091204583645, + 0.021235063672065735, + 0.008502142503857613, + 0.012341812252998352, + 0.008790123276412487, + 0.007233297452330589, + 0.011636510491371155, + 0.009255567565560341, + 0.005550004076212645, + 0.012852507643401623, + 0.005851017311215401, + 0.00578304473310709, + 0.00838795118033886, + 0.008494439534842968, + 0.014394915662705898, + 0.013055630028247833, + 0.007698375266045332, + 0.023145459592342377, + 0.003810881869867444, + 0.004960191436111927, + 0.004691760521382093, + 0.004452855791896582, + 0.005847888067364693, + 0.0047968123108148575, + 0.00395330460742116, + 0.0055225142277777195, + 0.0226057730615139, + 0.019952600821852684, + 0.04051130637526512, + 0.012081771157681942, + 0.012941981665790081, + 0.015873301774263382, + 0.01008022204041481, + 0.008466627448797226, + 0.009536098688840866, + 0.010500792413949966, + 0.013485544361174107, + 0.0076189106330275536, + 0.010471470654010773, + 0.011359918862581253, + 0.009950421750545502, + 0.015558362007141113, + 0.02033005841076374, + 0.016840055584907532, + 0.010541318915784359, + 0.006377596873790026, + 0.007015665527433157, + 0.010572143830358982, + 0.007690535392612219, + 0.01679329201579094, + 0.010882146656513214, + 0.0077004386112093925, + 0.009880749508738518, + 0.005679968744516373, + 0.008199247531592846, + 0.008914235979318619, + 0.007161639165133238, + 0.010301949456334114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "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 shot a static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609749481081963 + ], + "opposite_logprob_tokens": [ + -3.8956098556518555 + ], + "logprob_reward": -0.020609749481081963, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609749481081963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23691222071647644, + 0.11337887495756149, + 0.07520481944084167, + 0.05126817524433136, + 0.048230577260255814, + 0.033523838967084885, + 0.05772503465414047, + 0.15097689628601074, + 0.0957716703414917, + 0.137007936835289 + ], + "credit_map": [ + 0.02803533710539341, + 0.027120519429445267, + 0.024318715557456017, + 0.015438234433531761, + 0.05897745117545128, + 0.03263756260275841, + 0.026889069005846977, + 0.02349531091749668, + 0.01165593508630991, + 0.008814753033220768, + 0.01177993044257164, + 0.006979318801313639, + 0.029647888615727425, + 0.016180722042918205, + 0.015575353987514973, + 0.012744968757033348, + 0.006656443700194359, + 0.006658039055764675, + 0.007583625614643097, + 0.004559707827866077, + 0.01994151435792446, + 0.010626832954585552, + 0.008649908937513828, + 0.010528739541769028, + 0.004911583382636309, + 0.006018282845616341, + 0.0050277188420295715, + 0.003724279347807169, + 0.011477264575660229, + 0.007460268680006266, + 0.0050661745481193066, + 0.007582597900182009, + 0.0037717814557254314, + 0.004965405911207199, + 0.004851502366364002, + 0.0032950262539088726, + 0.008877820335328579, + 0.005982288159430027, + 0.0070572360418736935, + 0.0094295097514987, + 0.002569411648437381, + 0.0028139452915638685, + 0.003482302650809288, + 0.0037564984522759914, + 0.008505875244736671, + 0.005273721646517515, + 0.003455784637480974, + 0.0036663005594164133, + 0.007343149743974209, + 0.006854574661701918, + 0.011840933933854103, + 0.004935969598591328, + 0.009379489347338676, + 0.007136845961213112, + 0.004922945983707905, + 0.005311119835823774, + 0.016817500814795494, + 0.016490206122398376, + 0.02524290233850479, + 0.011680088005959988, + 0.027567582204937935, + 0.018877282738685608, + 0.01610778644680977, + 0.018193531781435013, + 0.010393360629677773, + 0.00847660843282938, + 0.014481201767921448, + 0.010503270663321018, + 0.020918739959597588, + 0.011156498454511166, + 0.009333294816315174, + 0.010508691892027855, + 0.021042972803115845, + 0.017438914626836777, + 0.016617896035313606, + 0.012696321122348309, + 0.01911837048828602, + 0.015193497762084007, + 0.012945529073476791, + 0.021954407915472984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.020609749481081963 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001341158407740295 + ], + "opposite_logprob_tokens": [ + -6.626341342926025 + ], + "logprob_reward": -0.001341158407740295, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001341158407740295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2336672842502594, + 0.0679454654455185, + 0.0422532819211483, + 0.049915626645088196, + 0.12142850458621979, + 0.030327387154102325, + 0.2111034095287323, + 0.09220738708972931, + 0.08465928584337234, + 0.06649236381053925 + ], + "credit_map": [ + 0.030474422499537468, + 0.03298626095056534, + 0.02392195165157318, + 0.013800498098134995, + 0.048259783536195755, + 0.03101603500545025, + 0.029602332040667534, + 0.023605985566973686, + 0.006998374126851559, + 0.0059198676608502865, + 0.006256196182221174, + 0.005224652122706175, + 0.01296934112906456, + 0.009260725229978561, + 0.01009878609329462, + 0.011217519640922546, + 0.004256191197782755, + 0.004686689469963312, + 0.004761917982250452, + 0.002647176617756486, + 0.006110050715506077, + 0.005368879996240139, + 0.0049330671317875385, + 0.009489305317401886, + 0.0047115725465118885, + 0.004346091765910387, + 0.0056333537213504314, + 0.004086358938366175, + 0.012656636536121368, + 0.00651925103738904, + 0.004279261454939842, + 0.007683099247515202, + 0.01566977985203266, + 0.019314339384436607, + 0.013703283853828907, + 0.01013244315981865, + 0.010466462932527065, + 0.014278119429945946, + 0.011927935294806957, + 0.02593613788485527, + 0.003181277774274349, + 0.0034259981475770473, + 0.002988709369674325, + 0.0027536542620509863, + 0.005577042233198881, + 0.004204403143376112, + 0.0033947222400456667, + 0.004801580216735601, + 0.03611786663532257, + 0.04159417003393173, + 0.032862041145563126, + 0.02144610695540905, + 0.025579312816262245, + 0.023288525640964508, + 0.014457683078944683, + 0.015757689252495766, + 0.011477374471724033, + 0.014009583741426468, + 0.010671097785234451, + 0.009729730896651745, + 0.010374854318797588, + 0.013064620085060596, + 0.007900871336460114, + 0.014979255385696888, + 0.014787509106099606, + 0.009982182644307613, + 0.013339366763830185, + 0.0073042381554841995, + 0.007493588142096996, + 0.008955146186053753, + 0.007461274042725563, + 0.015335981734097004, + 0.013048275373876095, + 0.009453188627958298, + 0.008037040941417217, + 0.005658671725541353, + 0.008277890272438526, + 0.008235099725425243, + 0.005760552361607552, + 0.008021640591323376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.001341158407740295 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627411279827356 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627411279827356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627411279827356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3372262120246887, + 0.07917206734418869, + 0.09058555960655212, + 0.04516051337122917, + 0.08154188841581345, + 0.06727824360132217, + 0.09531613439321518, + 0.0553225576877594, + 0.10899742692708969, + 0.0393994003534317 + ], + "credit_map": [ + 0.03810219466686249, + 0.039116065949201584, + 0.06408735364675522, + 0.030502503737807274, + 0.04447716474533081, + 0.041553061455488205, + 0.03301190584897995, + 0.04637596383690834, + 0.009050270542502403, + 0.007845869287848473, + 0.008948525413870811, + 0.005534717813134193, + 0.012784061953425407, + 0.01218147948384285, + 0.00814505759626627, + 0.014682088978588581, + 0.008808411657810211, + 0.008146067149937153, + 0.009205104783177376, + 0.005273872520774603, + 0.010482709854841232, + 0.012065448798239231, + 0.009714984335005283, + 0.02688896656036377, + 0.004800604656338692, + 0.006030229385942221, + 0.005462199449539185, + 0.004155394155532122, + 0.005654693581163883, + 0.005562599282711744, + 0.003919313661754131, + 0.00957548338919878, + 0.008714529685676098, + 0.010345308110117912, + 0.008679910562932491, + 0.005911612883210182, + 0.0069044409319758415, + 0.010433993302285671, + 0.009246713481843472, + 0.02130538411438465, + 0.004585533868521452, + 0.006502869073301554, + 0.011699535883963108, + 0.01591748744249344, + 0.004233859945088625, + 0.006627046503126621, + 0.006273277103900909, + 0.011438639834523201, + 0.01748194731771946, + 0.018882131204009056, + 0.013121683150529861, + 0.009281938895583153, + 0.012899693101644516, + 0.010777625255286694, + 0.006652672775089741, + 0.006218457594513893, + 0.00733003206551075, + 0.010785055346786976, + 0.0057089547626674175, + 0.00521067064255476, + 0.006657843478024006, + 0.007940851151943207, + 0.005124547053128481, + 0.006564608309417963, + 0.02540011890232563, + 0.01828843355178833, + 0.013059856370091438, + 0.006806548219174147, + 0.006798053625971079, + 0.012016376480460167, + 0.007563741412013769, + 0.01906430348753929, + 0.005055885761976242, + 0.003735918551683426, + 0.006533222738653421, + 0.0035076464992016554, + 0.00449354387819767, + 0.004002641886472702, + 0.004159606993198395, + 0.007910938933491707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.007627411279827356 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252846051007509 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.005252846051007509, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252846051007509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12361583113670349, + 0.05790716037154198, + 0.10205405205488205, + 0.0673578679561615, + 0.24817882478237152, + 0.019911490380764008, + 0.08969602733850479, + 0.06074317917227745, + 0.03135974332690239, + 0.1991758495569229 + ], + "credit_map": [ + 0.013518060557544231, + 0.012596038170158863, + 0.015139427036046982, + 0.007590885739773512, + 0.02498314529657364, + 0.016648102551698685, + 0.01906142197549343, + 0.01407875306904316, + 0.007219661492854357, + 0.008551288396120071, + 0.009684714488685131, + 0.006428856402635574, + 0.006834068801254034, + 0.0066498820669949055, + 0.005990047939121723, + 0.006548635195940733, + 0.008904599584639072, + 0.009722079150378704, + 0.016048133373260498, + 0.006933064199984074, + 0.009932770393788815, + 0.013119549490511417, + 0.011766057461500168, + 0.025627801194787025, + 0.0091178547590971, + 0.010930236428976059, + 0.008715374395251274, + 0.006998438388109207, + 0.006838821340352297, + 0.007662063464522362, + 0.007304272148758173, + 0.009790808893740177, + 0.03825198486447334, + 0.04910087585449219, + 0.02879706770181656, + 0.027754688635468483, + 0.02215423248708248, + 0.030817989259958267, + 0.015116583555936813, + 0.03618539124727249, + 0.0020752414129674435, + 0.0022991555742919445, + 0.0022432825062423944, + 0.002840605331584811, + 0.0020842556841671467, + 0.0022244001738727093, + 0.001920142094604671, + 0.004224407486617565, + 0.015344025567173958, + 0.017225399613380432, + 0.012816048227250576, + 0.009094267152249813, + 0.011125036515295506, + 0.009998264722526073, + 0.006545204669237137, + 0.0075477794744074345, + 0.00799400731921196, + 0.008996549062430859, + 0.007856911979615688, + 0.0052599310874938965, + 0.006027924362570047, + 0.007422839291393757, + 0.004730944987386465, + 0.012454072944819927, + 0.005100449081510305, + 0.004456149414181709, + 0.0032622467260807753, + 0.0025837337598204613, + 0.0038873793091624975, + 0.004227002151310444, + 0.003226139582693577, + 0.004616643767803907, + 0.04405248165130615, + 0.0246554184705019, + 0.03531460091471672, + 0.015291715040802956, + 0.010635411366820335, + 0.020177368074655533, + 0.014005539938807487, + 0.03504330664873123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.005252846051007509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, rendered with a clean digital aesthetic, captures a poignant nighttime scene. Yuzuru Nishimiya, a young girl with short, light brown hair gently framing her face, stands in the center of a simple concrete pedestrian bridge. She wears a classic Japanese school uniform, consisting of a white, sailor-style shirt and a dark, pleated skirt, the fabric appearing almost black in the low light. The bright, cool glow of a large, full moon, just rising above the horizon, casts her figure into a stark, defined silhouette, while depicting delicate silver rims of light along her hair and shoulders. Below the bridge, a slender reflection of the moon shimmers on the deep, dark waters of a slow-flowing river. Her pose is one of profound stillness, her hands resting lightly on the cool metal railing, her gaze downward. Her facial expression is lost in the shadows, yet her body conveys a deep sense of quiet contemplation. The cool evening air is noticeably crisp, almost tangible, and in the distance the soft glow of city lights dissolves into a beautiful, softly warm-toned bokeh, creating a serene and melancholic atmosphere that isolates her in this introspective moment.", + "seed": 8802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8802_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": -0.016560749756172298, + "vlm_logprob": -0.016560749756172298, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.016560749756172298, + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023689333349466324 + ], + "opposite_logprob_tokens": [ + -4.148689270019531 + ], + "logprob_reward": -0.023689333349466324, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023689333349466324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3281615376472473, + 0.09752196818590164, + 0.11571372300386429, + 0.0444980151951313, + 0.038019001483917236, + 0.06811738759279251, + 0.14312328398227692, + 0.10287223756313324, + 0.02428768202662468, + 0.03768518194556236 + ], + "credit_map": [ + 0.07327909767627716, + 0.03782976046204567, + 0.041373010724782944, + 0.03179345652461052, + 0.03837875649333, + 0.04370156303048134, + 0.04010026156902313, + 0.021705618128180504, + 0.02304602600634098, + 0.01066830288618803, + 0.011413268744945526, + 0.009842929430305958, + 0.010499930940568447, + 0.01337353978306055, + 0.011762294918298721, + 0.006915677804499865, + 0.02546187862753868, + 0.013882973231375217, + 0.013765849173069, + 0.010852036066353321, + 0.01498466171324253, + 0.016479380428791046, + 0.014250915497541428, + 0.006036026868969202, + 0.008960673585534096, + 0.004615511279553175, + 0.0058275628834962845, + 0.004948402754962444, + 0.005328092724084854, + 0.0056165726855397224, + 0.004546032287180424, + 0.004655166529119015, + 0.006144501268863678, + 0.006262503564357758, + 0.005348054692149162, + 0.0033398231025785208, + 0.005212295800447464, + 0.004911316093057394, + 0.0033522218000143766, + 0.003448287956416607, + 0.010059346444904804, + 0.008915385231375694, + 0.00827089138329029, + 0.007012787740677595, + 0.011265145614743233, + 0.008955370634794235, + 0.007325656712055206, + 0.006312801036983728, + 0.023087747395038605, + 0.018220948055386543, + 0.014980020001530647, + 0.016672682017087936, + 0.01908920705318451, + 0.019735917448997498, + 0.01834968477487564, + 0.012987068854272366, + 0.014174244366586208, + 0.01177811250090599, + 0.009489991702139378, + 0.014602825976908207, + 0.014238270930945873, + 0.016482338309288025, + 0.012522506527602673, + 0.009583947248756886, + 0.005204434972256422, + 0.003331472398713231, + 0.003207411617040634, + 0.0022721062414348125, + 0.0025117918848991394, + 0.003505233209580183, + 0.0025900339242070913, + 0.0016651964979246259, + 0.006003594025969505, + 0.005809227004647255, + 0.003930568229407072, + 0.004812413826584816, + 0.00471343332901597, + 0.004850306548178196, + 0.004373850300908089, + 0.0031917865853756666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.023689333349466324 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4465962052345276, + 0.15700571238994598, + 0.08971524983644485, + 0.0301384087651968, + 0.06165463104844093, + 0.044853582978248596, + 0.045545417815446854, + 0.04915513098239899, + 0.04169154539704323, + 0.033644091337919235 + ], + "credit_map": [ + 0.07429324090480804, + 0.05095445737242699, + 0.041978009045124054, + 0.05948755145072937, + 0.07054230570793152, + 0.05509879067540169, + 0.06951256096363068, + 0.024729298427700996, + 0.02574753388762474, + 0.017031680792570114, + 0.016353687271475792, + 0.02063533291220665, + 0.0224821325391531, + 0.018726704642176628, + 0.02572040632367134, + 0.010308233089745045, + 0.012268932536244392, + 0.010121314786374569, + 0.00807238556444645, + 0.012816959992051125, + 0.012638074345886707, + 0.011078636161983013, + 0.017278335988521576, + 0.005440606735646725, + 0.005761789157986641, + 0.0037436503916978836, + 0.004223119933158159, + 0.003388495882973075, + 0.003521349048241973, + 0.004090347792953253, + 0.0034732779022306204, + 0.0019363772589713335, + 0.01088299136608839, + 0.008152213878929615, + 0.006620058789849281, + 0.007278874982148409, + 0.008304495364427567, + 0.007751771714538336, + 0.00854066014289856, + 0.004123561549931765, + 0.007778380066156387, + 0.005251440219581127, + 0.005844730418175459, + 0.005172933451831341, + 0.005478175822645426, + 0.005655244924128056, + 0.005698691122233868, + 0.003973989747464657, + 0.009266246110200882, + 0.006190780084580183, + 0.0047267707996070385, + 0.004102720879018307, + 0.006345843430608511, + 0.006278671324253082, + 0.004934708587825298, + 0.0036996735725551844, + 0.008301425725221634, + 0.005747335497289896, + 0.005400404799729586, + 0.005629592575132847, + 0.00697278743609786, + 0.006308626849204302, + 0.006927591748535633, + 0.003867365885525942, + 0.008209201507270336, + 0.006017144303768873, + 0.0043923575431108475, + 0.003938613459467888, + 0.006248504389077425, + 0.006355051416903734, + 0.0040497141890227795, + 0.0024809595197439194, + 0.0063976626843214035, + 0.004780634772032499, + 0.004543195012956858, + 0.003450461197644472, + 0.003898377064615488, + 0.004578232765197754, + 0.0037082727067172527, + 0.002287256997078657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02637290023267269 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3443971276283264, + 0.14202116429805756, + 0.18678398430347443, + 0.0490189827978611, + 0.05959494411945343, + 0.04948551207780838, + 0.042207881808280945, + 0.023910129442811012, + 0.04103478789329529, + 0.06154541298747063 + ], + "credit_map": [ + 0.07372483611106873, + 0.04058701917529106, + 0.044929057359695435, + 0.028092270717024803, + 0.037473227828741074, + 0.0453575998544693, + 0.05169747397303581, + 0.02253567986190319, + 0.026463355869054794, + 0.02106473594903946, + 0.015181283466517925, + 0.013689713552594185, + 0.017033323645591736, + 0.02072743885219097, + 0.017956115305423737, + 0.009905216284096241, + 0.038287203758955, + 0.02722139284014702, + 0.02332853339612484, + 0.015627972781658173, + 0.02386324666440487, + 0.030116334557533264, + 0.018175296485424042, + 0.010164018720388412, + 0.009002511389553547, + 0.00811096653342247, + 0.004897011443972588, + 0.0042578838765621185, + 0.005515596829354763, + 0.006564151495695114, + 0.0068912506103515625, + 0.0037796187680214643, + 0.011262502521276474, + 0.006506912410259247, + 0.006906488910317421, + 0.006078863982111216, + 0.008313956670463085, + 0.008198381401598454, + 0.007447558920830488, + 0.004880285356193781, + 0.008140869438648224, + 0.0059461575001478195, + 0.005557751748710871, + 0.005353965796530247, + 0.00825803168118, + 0.006955764256417751, + 0.0064563085325062275, + 0.0028166717384010553, + 0.008374745957553387, + 0.005050197709351778, + 0.005150581244379282, + 0.0033884954173117876, + 0.005269723944365978, + 0.005444692447781563, + 0.006215787027031183, + 0.003313664346933365, + 0.004657347686588764, + 0.003349937731400132, + 0.0023699121084064245, + 0.0025260155089199543, + 0.003063003532588482, + 0.003248737659305334, + 0.0030852381605654955, + 0.0016099408967420459, + 0.005951223894953728, + 0.005806528031826019, + 0.0038888133130967617, + 0.00531832966953516, + 0.00650783721357584, + 0.005986358970403671, + 0.0048612793907523155, + 0.002714421832934022, + 0.01164779532700777, + 0.010847561061382294, + 0.0054662227630615234, + 0.00626846868544817, + 0.009253699332475662, + 0.00884146336466074, + 0.005908478982746601, + 0.003311732318252325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061680391430855 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061680391430855, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061680391430855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4524237811565399, + 0.1116611659526825, + 0.10688918828964233, + 0.052502717822790146, + 0.04975735768675804, + 0.05755539610981941, + 0.03582916781306267, + 0.04845782741904259, + 0.03842131420969963, + 0.046502113342285156 + ], + "credit_map": [ + 0.09127509593963623, + 0.04257941618561745, + 0.08008238673210144, + 0.033917348831892014, + 0.04557524621486664, + 0.07832735031843185, + 0.05108345299959183, + 0.02958344668149948, + 0.020702172070741653, + 0.00965458806604147, + 0.017964515835046768, + 0.008713282644748688, + 0.010753434151411057, + 0.021752353757619858, + 0.01092502661049366, + 0.011195795610547066, + 0.02256927266716957, + 0.009811765514314175, + 0.0232455562800169, + 0.0073599196039140224, + 0.009805758483707905, + 0.01874273270368576, + 0.008356790989637375, + 0.006997392512857914, + 0.012131774798035622, + 0.004935527220368385, + 0.009450145065784454, + 0.003581835888326168, + 0.0049179368652403355, + 0.008758565410971642, + 0.004328813403844833, + 0.004398119170218706, + 0.010194290429353714, + 0.005037068855017424, + 0.009564374573528767, + 0.0035336080472916365, + 0.0057801115326583385, + 0.007847964763641357, + 0.00457358593121171, + 0.0032263528555631638, + 0.0123475706204772, + 0.005692411679774523, + 0.009345904923975468, + 0.0035160758998245, + 0.006699315272271633, + 0.010325068607926369, + 0.005458869971334934, + 0.004170177038758993, + 0.006843544542789459, + 0.0032654828391969204, + 0.006584540940821171, + 0.0025447190273553133, + 0.004523442126810551, + 0.006291599478572607, + 0.003677806118503213, + 0.0020980339031666517, + 0.007546177599579096, + 0.005801815539598465, + 0.004786301404237747, + 0.005157764535397291, + 0.006535529624670744, + 0.007839455269277096, + 0.006447962485253811, + 0.004342825151979923, + 0.006728326436132193, + 0.004419391509145498, + 0.0060256654396653175, + 0.004188595339655876, + 0.004769605118781328, + 0.005598006770014763, + 0.00431467778980732, + 0.0023770432453602552, + 0.007693642750382423, + 0.005019286647439003, + 0.007071294821798801, + 0.0035094022750854492, + 0.006247816141694784, + 0.007689496036618948, + 0.005512693431228399, + 0.003758479608222842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016061680391430855 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188248664140701 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188248664140701, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188248664140701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45454585552215576, + 0.11161220818758011, + 0.08441353589296341, + 0.04693435877561569, + 0.03250117599964142, + 0.08173421025276184, + 0.050223544239997864, + 0.057521332055330276, + 0.03701474890112877, + 0.043499115854501724 + ], + "credit_map": [ + 0.11310447752475739, + 0.04142942652106285, + 0.06314542889595032, + 0.03182235732674599, + 0.045654669404029846, + 0.07077188789844513, + 0.05488035827875137, + 0.03373723104596138, + 0.022917702794075012, + 0.012986728921532631, + 0.013013884425163269, + 0.008019329980015755, + 0.012821860611438751, + 0.01970183104276657, + 0.0092967189848423, + 0.012854146771132946, + 0.016291387379169464, + 0.009113588370382786, + 0.017366014420986176, + 0.00547712342813611, + 0.0076593076810240746, + 0.014692862518131733, + 0.007830142974853516, + 0.005983111914247274, + 0.011871442198753357, + 0.005695107392966747, + 0.005718734115362167, + 0.003337886882945895, + 0.004605287220329046, + 0.007860411889851093, + 0.0039197844453155994, + 0.003925701137632132, + 0.005218901205807924, + 0.003226183820515871, + 0.0051992605440318584, + 0.003024848410859704, + 0.0038397458847612143, + 0.005097488407045603, + 0.003990426193922758, + 0.002904319204390049, + 0.01540739368647337, + 0.010014036670327187, + 0.009979085065424442, + 0.005746881943196058, + 0.010465181432664394, + 0.012365087866783142, + 0.01073045190423727, + 0.007026086561381817, + 0.008817678317427635, + 0.00546177988871932, + 0.0076153394766151905, + 0.003470863215625286, + 0.0064314668998122215, + 0.007368259131908417, + 0.006690744776278734, + 0.0043674129992723465, + 0.009588339366018772, + 0.006995150353759527, + 0.006665871478617191, + 0.00513225793838501, + 0.0074213300831615925, + 0.008017820306122303, + 0.008093046024441719, + 0.005607516970485449, + 0.007348116021603346, + 0.003605426289141178, + 0.006038289051502943, + 0.0030695945024490356, + 0.004488379694521427, + 0.005989963188767433, + 0.0034565171226859093, + 0.0030184625647962093, + 0.006004849448800087, + 0.004904024302959442, + 0.005184590350836515, + 0.0041353474371135235, + 0.006115004885941744, + 0.006383304484188557, + 0.006622422020882368, + 0.004149573389440775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014188248664140701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, rendered with a clean digital aesthetic, captures a poignant nighttime scene. Yuzuru Nishimiya, a young girl with short, light brown hair gently framing her face, stands in the center of a simple concrete pedestrian bridge. She wears a classic Japanese school uniform, consisting of a white, sailor-style shirt and a dark, pleated skirt, the fabric appearing almost black in the low light. The bright, cool glow of a large, full moon, just rising above the horizon, casts her figure into a stark, defined silhouette, while depicting delicate silver rims of light along her hair and shoulders. Below the bridge, a slender reflection of the moon shimmers on the deep, dark waters of a slow-flowing river. Her pose is one of profound stillness, her hands resting lightly on the cool metal railing, her gaze downward. Her facial expression is lost in the shadows, yet her body conveys a deep sense of quiet contemplation. The cool evening air is noticeably crisp, almost tangible, and in the distance the soft glow of city lights dissolves into a beautiful, softly warm-toned bokeh, creating a serene and melancholic atmosphere that isolates her in this introspective moment.", + "seed": 8803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8803_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": -0.010689955833368004, + "vlm_logprob": -0.010689955833368004, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.010689955833368004, + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015725649893283844 + ], + "opposite_logprob_tokens": [ + -4.515725612640381 + ], + "logprob_reward": -0.015725649893283844, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015725649893283844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2382265031337738, + 0.11184794455766678, + 0.07583130896091461, + 0.07533392310142517, + 0.14386342465877533, + 0.0683899074792862, + 0.04017971456050873, + 0.08841688185930252, + 0.06366156786680222, + 0.09424879401922226 + ], + "credit_map": [ + 0.0290298480540514, + 0.023442627862095833, + 0.035694319754838943, + 0.012617294676601887, + 0.05174744129180908, + 0.027839938178658485, + 0.03732355684041977, + 0.020531466230750084, + 0.013883360661566257, + 0.010288573801517487, + 0.011910663917660713, + 0.007361254654824734, + 0.025942666456103325, + 0.015143493190407753, + 0.0177619531750679, + 0.009555978700518608, + 0.01056445948779583, + 0.009689138270914555, + 0.010904526337981224, + 0.0042001428082585335, + 0.013819531537592411, + 0.007545103318989277, + 0.012464373372495174, + 0.006644035689532757, + 0.006419188342988491, + 0.007169324439018965, + 0.00992392934858799, + 0.00574942771345377, + 0.012237095274031162, + 0.009389468468725681, + 0.01563427783548832, + 0.008811211213469505, + 0.011273493058979511, + 0.010493103414773941, + 0.017252812162041664, + 0.009980128146708012, + 0.03196873143315315, + 0.020670462399721146, + 0.03043227270245552, + 0.011792420409619808, + 0.007940519601106644, + 0.007430596277117729, + 0.008048619143664837, + 0.004609682597219944, + 0.014538601972162724, + 0.007758685387670994, + 0.010924902744591236, + 0.007138299290090799, + 0.005743144545704126, + 0.003582477569580078, + 0.006166127044707537, + 0.0020959496032446623, + 0.007240646984428167, + 0.0049658105708658695, + 0.006831607315689325, + 0.003553952556103468, + 0.007055502850562334, + 0.006465727463364601, + 0.010878403671085835, + 0.00719488924369216, + 0.01697005145251751, + 0.011096889153122902, + 0.018434496596455574, + 0.010320920497179031, + 0.005978632252663374, + 0.006048913113772869, + 0.007379116490483284, + 0.005252030678093433, + 0.012596076354384422, + 0.007430944126099348, + 0.010605134069919586, + 0.008370720781385899, + 0.008833029307425022, + 0.007276986259967089, + 0.01018899492919445, + 0.0072986166924238205, + 0.02255968376994133, + 0.013310441747307777, + 0.01722503826022148, + 0.007555999327450991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.015725649893283844 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821165345609188 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.01821165345609188, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821165345609188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23011769354343414, + 0.20192137360572815, + 0.05609303340315819, + 0.05384676903486252, + 0.17571371793746948, + 0.061453238129615784, + 0.034945111721754074, + 0.030468866229057312, + 0.024867748841643333, + 0.13057242333889008 + ], + "credit_map": [ + 0.018381807953119278, + 0.01973458006978035, + 0.027772951871156693, + 0.016976822167634964, + 0.04646623134613037, + 0.0316547155380249, + 0.03982818126678467, + 0.029302416369318962, + 0.015071568079292774, + 0.014141105115413666, + 0.022866714745759964, + 0.013797069899737835, + 0.04145617038011551, + 0.02935836650431156, + 0.04629902169108391, + 0.018931377679109573, + 0.004792360123246908, + 0.005162726156413555, + 0.008979576639831066, + 0.0035615689121186733, + 0.008304635062813759, + 0.006703572813421488, + 0.0101009551435709, + 0.008487638086080551, + 0.005967426113784313, + 0.005563021637499332, + 0.010114399716258049, + 0.0049965339712798595, + 0.007436934858560562, + 0.006115077529102564, + 0.005709892604500055, + 0.007943489588797092, + 0.013552599586546421, + 0.01266754511743784, + 0.01897663064301014, + 0.011999819427728653, + 0.03605746105313301, + 0.02585083618760109, + 0.04014548659324646, + 0.01646333932876587, + 0.00815639179199934, + 0.0067766522988677025, + 0.007511925883591175, + 0.004006919451057911, + 0.012644470669329166, + 0.007634145673364401, + 0.009747196920216084, + 0.00497553963214159, + 0.0029475155752152205, + 0.00244156364351511, + 0.003998954314738512, + 0.0028205099515616894, + 0.006743622478097677, + 0.004332252312451601, + 0.007754356134682894, + 0.003906338009983301, + 0.0032029214780777693, + 0.002445128746330738, + 0.0048619103617966175, + 0.00249481201171875, + 0.004388852510601282, + 0.0036309652496129274, + 0.005441024899482727, + 0.004003251437097788, + 0.002868546172976494, + 0.0017687906511127949, + 0.004109995439648628, + 0.0018524689367040992, + 0.0038711377419531345, + 0.0030984424520283937, + 0.003857177682220936, + 0.00344119337387383, + 0.010689569637179375, + 0.00905191246420145, + 0.01609315536916256, + 0.009169971570372581, + 0.022299272939562798, + 0.017187178134918213, + 0.03263106942176819, + 0.01345030777156353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01821165345609188 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017137139802798629 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017137139802798629, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017137139802798629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28500330448150635, + 0.14736579358577728, + 0.13126429915428162, + 0.04929045960307121, + 0.14963184297084808, + 0.03884539380669594, + 0.03533367067575455, + 0.061269573867321014, + 0.032302577048540115, + 0.06969312578439713 + ], + "credit_map": [ + 0.03317570686340332, + 0.02833005040884018, + 0.04142981395125389, + 0.01624682918190956, + 0.05222051963210106, + 0.035857729613780975, + 0.049228738993406296, + 0.028513871133327484, + 0.018390921875834465, + 0.015700243413448334, + 0.014568948186933994, + 0.009958993643522263, + 0.029811080545186996, + 0.02095983549952507, + 0.026626085862517357, + 0.011349671520292759, + 0.014385608024895191, + 0.012059212662279606, + 0.015754204243421555, + 0.008029945194721222, + 0.027920057997107506, + 0.01923723705112934, + 0.0240347720682621, + 0.00984326470643282, + 0.005840410944074392, + 0.005572707392275333, + 0.006388609763234854, + 0.003259415505453944, + 0.009046927094459534, + 0.0074669294990599155, + 0.007980830036103725, + 0.0037346286699175835, + 0.01300840824842453, + 0.009899236261844635, + 0.020818011835217476, + 0.009492586366832256, + 0.030362386256456375, + 0.022830061614513397, + 0.03147345781326294, + 0.011747690849006176, + 0.004130198620259762, + 0.003374045714735985, + 0.00426413444802165, + 0.002562823472544551, + 0.00810540746897459, + 0.005581903271377087, + 0.007111093029379845, + 0.0037157845217734575, + 0.0049660345539450645, + 0.0031499676406383514, + 0.005510467570275068, + 0.0019002313492819667, + 0.0065550897270441055, + 0.00521079683676362, + 0.005281051155179739, + 0.00276002986356616, + 0.005885428749024868, + 0.00554141728207469, + 0.0073090107180178165, + 0.003982607275247574, + 0.0118593480437994, + 0.008622964844107628, + 0.012038382701575756, + 0.006030410062521696, + 0.0039937240071594715, + 0.0036560960579663515, + 0.005191703327000141, + 0.0017430493608117104, + 0.005378860514611006, + 0.003824666142463684, + 0.003794901305809617, + 0.0047195772640407085, + 0.006738162133842707, + 0.0053647831082344055, + 0.00903172418475151, + 0.004720773082226515, + 0.016625503078103065, + 0.011327741667628288, + 0.011491104029119015, + 0.004393326584249735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0017137139802798629 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730386987328529 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730386987328529, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730386987328529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33756861090660095, + 0.14909346401691437, + 0.08167842775583267, + 0.14912353456020355, + 0.09720426797866821, + 0.040179211646318436, + 0.042077019810676575, + 0.03139466792345047, + 0.024679120630025864, + 0.047001611441373825 + ], + "credit_map": [ + 0.04115758091211319, + 0.030927682295441628, + 0.06845366954803467, + 0.026979153975844383, + 0.039995018392801285, + 0.03866039589047432, + 0.0457855649292469, + 0.04560955613851547, + 0.020286070182919502, + 0.013378461822867393, + 0.028445009142160416, + 0.012270202860236168, + 0.018626688048243523, + 0.01821330189704895, + 0.02045731246471405, + 0.01741640828549862, + 0.011066233739256859, + 0.007564779836684465, + 0.015246661379933357, + 0.005785988178104162, + 0.010476024821400642, + 0.009508087299764156, + 0.011776546016335487, + 0.010254107415676117, + 0.02105720527470112, + 0.014151412062346935, + 0.027338046580553055, + 0.009971442632377148, + 0.021456383168697357, + 0.0194053053855896, + 0.021207714453339577, + 0.01453602034598589, + 0.007640279829502106, + 0.006708128843456507, + 0.01406550221145153, + 0.007614292204380035, + 0.01448568794876337, + 0.012127920053899288, + 0.0231462549418211, + 0.011416194960474968, + 0.005498949903994799, + 0.0039035743102431297, + 0.006919561419636011, + 0.0028490310069173574, + 0.006477131973952055, + 0.00505384337157011, + 0.005224539432674646, + 0.004252582788467407, + 0.0069687627255916595, + 0.003464676672592759, + 0.007331832777708769, + 0.0029214308597147465, + 0.006644320674240589, + 0.005672271829098463, + 0.005327135324478149, + 0.0037465908098965883, + 0.0029836224857717752, + 0.002642466453835368, + 0.004608769901096821, + 0.002762333955615759, + 0.005094877444207668, + 0.004049632232636213, + 0.0044701616279780865, + 0.004782801494002342, + 0.00264403922483325, + 0.002645699307322502, + 0.0033615906722843647, + 0.0019842826295644045, + 0.004885142669081688, + 0.0033912737853825092, + 0.003550879890099168, + 0.0022162122186273336, + 0.0055469935759902, + 0.003049676539376378, + 0.008664298802614212, + 0.002941740211099386, + 0.009589164517819881, + 0.007402017712593079, + 0.005994061008095741, + 0.003813661402091384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006730386987328529 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110683748498559 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.0110683748498559, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110683748498559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29082751274108887, + 0.14547619223594666, + 0.08535536378622055, + 0.11013565957546234, + 0.07510561496019363, + 0.025924762710928917, + 0.08195379376411438, + 0.08281783759593964, + 0.03715280443429947, + 0.06525042653083801 + ], + "credit_map": [ + 0.02975657768547535, + 0.025119202211499214, + 0.034625642001628876, + 0.02093077078461647, + 0.06494085490703583, + 0.0371120423078537, + 0.04738548770546913, + 0.030956929549574852, + 0.014960451982915401, + 0.010607342235744, + 0.016690703108906746, + 0.011365849524736404, + 0.021541450172662735, + 0.016503851860761642, + 0.024737387895584106, + 0.02906915545463562, + 0.009711389429867268, + 0.007321033626794815, + 0.009948601014912128, + 0.00602935254573822, + 0.012739789672195911, + 0.009244788438081741, + 0.016223164275288582, + 0.014137247577309608, + 0.01221143826842308, + 0.011137151159346104, + 0.014271372929215431, + 0.007360624149441719, + 0.016177434474229813, + 0.014660999178886414, + 0.02187737636268139, + 0.012439260259270668, + 0.009025473147630692, + 0.006839389447122812, + 0.009720347821712494, + 0.005298562813550234, + 0.010687178932130337, + 0.00793920923024416, + 0.012369305826723576, + 0.013226148672401905, + 0.0035003130324184895, + 0.0018364316783845425, + 0.0036773746833205223, + 0.0017397879855707288, + 0.004072088748216629, + 0.0034083460923284292, + 0.003321598516777158, + 0.004368821159005165, + 0.007258565630763769, + 0.005600478500127792, + 0.0075019532814621925, + 0.006548574659973383, + 0.015145761892199516, + 0.011150406673550606, + 0.018296552821993828, + 0.01045149751007557, + 0.008041794411838055, + 0.006737317889928818, + 0.007863631471991539, + 0.007343188859522343, + 0.015474582090973854, + 0.010941863059997559, + 0.017636224627494812, + 0.00877922773361206, + 0.004312853328883648, + 0.004815422464162111, + 0.004780429881066084, + 0.0025884381029754877, + 0.006839077919721603, + 0.004302246030420065, + 0.005764960311353207, + 0.003749377094209194, + 0.00684727868065238, + 0.004684379789978266, + 0.007750486955046654, + 0.004623448010534048, + 0.012099380604922771, + 0.009840762242674828, + 0.012122568674385548, + 0.007282121106982231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0110683748498559 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot, rendered with a clean digital aesthetic, captures a poignant nighttime scene. Yuzuru Nishimiya, a young girl with short, light brown hair gently framing her face, stands in the center of a simple concrete pedestrian bridge. She wears a classic Japanese school uniform, consisting of a white, sailor-style shirt and a dark, pleated skirt, the fabric appearing almost black in the low light. The bright, cool glow of a large, full moon, just rising above the horizon, casts her figure into a stark, defined silhouette, while depicting delicate silver rims of light along her hair and shoulders. Below the bridge, a slender reflection of the moon shimmers on the deep, dark waters of a slow-flowing river. Her pose is one of profound stillness, her hands resting lightly on the cool metal railing, her gaze downward. Her facial expression is lost in the shadows, yet her body conveys a deep sense of quiet contemplation. The cool evening air is noticeably crisp, almost tangible, and in the distance the soft glow of city lights dissolves into a beautiful, softly warm-toned bokeh, creating a serene and melancholic atmosphere that isolates her in this introspective moment.", + "seed": 8804, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_3/index_9680_seed_8804_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": -0.013220492540858686, + "vlm_logprob": -0.013220492540858686, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.013220492540858686, + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01290623378008604 + ], + "opposite_logprob_tokens": [ + -4.887906074523926 + ], + "logprob_reward": -0.01290623378008604, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01290623378008604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1724788248538971, + 0.08736888319253922, + 0.17134304344654083, + 0.0937509685754776, + 0.10290319472551346, + 0.04063555970788002, + 0.09332243353128433, + 0.030776740983128548, + 0.1064705029129982, + 0.10094983875751495 + ], + "credit_map": [ + 0.027505911886692047, + 0.022307032719254494, + 0.02198041044175625, + 0.01728120446205139, + 0.023026077076792717, + 0.031695518642663956, + 0.014369257725775242, + 0.014313410967588425, + 0.01010490395128727, + 0.01028991024941206, + 0.009502404369413853, + 0.0076437415555119514, + 0.011899915523827076, + 0.011113990098237991, + 0.011259665712714195, + 0.015554355457425117, + 0.023905429989099503, + 0.013460327871143818, + 0.020668452605605125, + 0.012626072391867638, + 0.01739373244345188, + 0.02424614131450653, + 0.022420626133680344, + 0.036622267216444016, + 0.012327641248703003, + 0.013188923709094524, + 0.009692435152828693, + 0.007752911187708378, + 0.010574234649538994, + 0.01425208244472742, + 0.011715728789567947, + 0.014247019775211811, + 0.01357852853834629, + 0.012332350946962833, + 0.011502474546432495, + 0.010944539681077003, + 0.010413376614451408, + 0.013631054200232029, + 0.011308418586850166, + 0.019192451611161232, + 0.006298774387687445, + 0.003984302282333374, + 0.0043053822591900826, + 0.0035255253314971924, + 0.005363842006772757, + 0.00613849051296711, + 0.0058503467589616776, + 0.00516889663413167, + 0.011965593323111534, + 0.009556083008646965, + 0.01185624860227108, + 0.007921711541712284, + 0.006530651822686195, + 0.012567521072924137, + 0.010225840844213963, + 0.022698787972331047, + 0.0030486928299069405, + 0.003174248617142439, + 0.004833587910979986, + 0.00268680602312088, + 0.0031475096475332975, + 0.004496206063777208, + 0.004103452432900667, + 0.0052862404845654964, + 0.011665788479149342, + 0.008932042866945267, + 0.010892397724092007, + 0.008884546346962452, + 0.00841184239834547, + 0.01470183115452528, + 0.018142186105251312, + 0.024839870631694794, + 0.011304927989840508, + 0.007331929635256529, + 0.01028906274586916, + 0.010649753734469414, + 0.0072370474226772785, + 0.013796193525195122, + 0.019014492630958557, + 0.02132643572986126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01290623378008604 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373682498931885 + ], + "opposite_logprob_tokens": [ + -3.4087367057800293 + ], + "logprob_reward": -0.03373682498931885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373682498931885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21027174592018127, + 0.03774799033999443, + 0.08968726545572281, + 0.07160671800374985, + 0.14412187039852142, + 0.230477437376976, + 0.04007913917303085, + 0.03383517265319824, + 0.05814450606703758, + 0.08402814716100693 + ], + "credit_map": [ + 0.02458667941391468, + 0.02062210440635681, + 0.02400493435561657, + 0.01870431751012802, + 0.03093891032040119, + 0.030411720275878906, + 0.026955416426062584, + 0.03404766321182251, + 0.0055946423672139645, + 0.004364266525954008, + 0.0037282027769833803, + 0.0037876241840422153, + 0.00513142254203558, + 0.00527358939871192, + 0.004631733521819115, + 0.005236506927758455, + 0.012128692120313644, + 0.006431202404201031, + 0.011235533282160759, + 0.00757027231156826, + 0.00950827170163393, + 0.011700885370373726, + 0.010023987852036953, + 0.021088428795337677, + 0.010450167581439018, + 0.005376165732741356, + 0.00900827907025814, + 0.005457235034555197, + 0.007519488222897053, + 0.013342299498617649, + 0.0070748222060501575, + 0.013378257863223553, + 0.021475009620189667, + 0.017301492393016815, + 0.015192131511867046, + 0.01299009658396244, + 0.0146165881305933, + 0.01640440709888935, + 0.01209518313407898, + 0.0340469665825367, + 0.030811376869678497, + 0.033908944576978683, + 0.0308507289737463, + 0.019948238506913185, + 0.027208635583519936, + 0.02877197414636612, + 0.017803646624088287, + 0.04117387905716896, + 0.0049005174078047276, + 0.0038641912396997213, + 0.00569854024797678, + 0.0036638795863837004, + 0.0029704375192523003, + 0.005366584286093712, + 0.005118850618600845, + 0.00849614292383194, + 0.004213464446365833, + 0.005077567417174578, + 0.0029289282392710447, + 0.0031592368613928556, + 0.00440280232578516, + 0.004511567298322916, + 0.003862492274492979, + 0.005679110996425152, + 0.00946244690567255, + 0.008697021752595901, + 0.00543052377179265, + 0.00494673615321517, + 0.005940662231296301, + 0.009757635183632374, + 0.006890691351145506, + 0.007018788252025843, + 0.011142284609377384, + 0.00868290662765503, + 0.00766695337370038, + 0.008588189259171486, + 0.007845455780625343, + 0.011996627785265446, + 0.014202308841049671, + 0.013903425075113773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03373682498931885 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010446811793372035 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010446811793372035, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010446811793372035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11335043609142303, + 0.05439789593219757, + 0.11400964856147766, + 0.04769938439130783, + 0.05409063771367073, + 0.10904410481452942, + 0.12095235288143158, + 0.15051928162574768, + 0.09664980322122574, + 0.13928645849227905 + ], + "credit_map": [ + 0.014036350883543491, + 0.012482549995183945, + 0.012144628912210464, + 0.008317914791405201, + 0.020402129739522934, + 0.015686292201280594, + 0.01355584617704153, + 0.0167247261852026, + 0.0058950092643499374, + 0.0070313578471541405, + 0.005975625943392515, + 0.004757288843393326, + 0.009867509827017784, + 0.006145056802779436, + 0.007285077590495348, + 0.00744097214192152, + 0.011812497861683369, + 0.0056727793999016285, + 0.015119576826691628, + 0.009428116492927074, + 0.00877787172794342, + 0.015046867541968822, + 0.018661752343177795, + 0.029490185901522636, + 0.00781924743205309, + 0.005568325985223055, + 0.003731972072273493, + 0.004279016051441431, + 0.006078608334064484, + 0.0063489326275885105, + 0.004682339262217283, + 0.009190940298140049, + 0.005516508594155312, + 0.004435809329152107, + 0.007001881953328848, + 0.006068852264434099, + 0.005033263470977545, + 0.007631100248545408, + 0.008976086974143982, + 0.009427130222320557, + 0.019428230822086334, + 0.013666106387972832, + 0.010986560955643654, + 0.009446496143937111, + 0.011227280832827091, + 0.018211334943771362, + 0.01078052818775177, + 0.015297573991119862, + 0.02168957330286503, + 0.0186328012496233, + 0.009112756699323654, + 0.011217953637242317, + 0.012042981572449207, + 0.017813028767704964, + 0.011643152683973312, + 0.01880009099841118, + 0.02071382664144039, + 0.025408191606402397, + 0.014600474387407303, + 0.014827108010649681, + 0.016641033813357353, + 0.019621850922703743, + 0.017841888591647148, + 0.02086488902568817, + 0.01044306717813015, + 0.01186842005699873, + 0.009921162389218807, + 0.008569731377065182, + 0.009808654896914959, + 0.012912098318338394, + 0.014905781485140324, + 0.01822088658809662, + 0.016534311696887016, + 0.010628074407577515, + 0.01499175000935793, + 0.01440533995628357, + 0.009700537659227848, + 0.018625279888510704, + 0.025739289820194244, + 0.02866186946630478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0010446811793372035 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0637529194355011, + 0.03947550803422928, + 0.039853256195783615, + 0.059612199664115906, + 0.22826266288757324, + 0.03571029752492905, + 0.04048047214746475, + 0.04322638735175133, + 0.15853700041770935, + 0.2910892367362976 + ], + "credit_map": [ + 0.009403812699019909, + 0.005599780473858118, + 0.00785328634083271, + 0.004803728777915239, + 0.009319732896983624, + 0.008744427934288979, + 0.006835037376731634, + 0.011193109676241875, + 0.004594975151121616, + 0.003074052045121789, + 0.004657556768506765, + 0.003372429171577096, + 0.006578715052455664, + 0.005355161149054766, + 0.004525411408394575, + 0.007317205425351858, + 0.0048983399756252766, + 0.00270366994664073, + 0.005562035832554102, + 0.0034056997392326593, + 0.003097319044172764, + 0.005098139401525259, + 0.004217134788632393, + 0.010870919562876225, + 0.007895154878497124, + 0.004959987010806799, + 0.008320137858390808, + 0.004813225939869881, + 0.006754247006028891, + 0.007141411304473877, + 0.006281785201281309, + 0.013446243479847908, + 0.0276695154607296, + 0.013144792057573795, + 0.03317601606249809, + 0.022763974964618683, + 0.017768017947673798, + 0.027765190228819847, + 0.0260150246322155, + 0.05996013060212135, + 0.004341833759099245, + 0.0035606424789875746, + 0.0037052701227366924, + 0.0027453701477497816, + 0.00493815541267395, + 0.0049672783352434635, + 0.0036093166563659906, + 0.007842428982257843, + 0.004265027120709419, + 0.0027563325129449368, + 0.007339814677834511, + 0.0028084286022931337, + 0.0030900766141712666, + 0.00589539110660553, + 0.0041107782162725925, + 0.010214623063802719, + 0.004952828399837017, + 0.0034180376678705215, + 0.004661696497350931, + 0.0033892362844198942, + 0.004506256431341171, + 0.006440471392124891, + 0.008461380377411842, + 0.007396481931209564, + 0.017564142122864723, + 0.013721324503421783, + 0.020805811509490013, + 0.014769412577152252, + 0.01357282791286707, + 0.02306482195854187, + 0.023312684148550034, + 0.03172598034143448, + 0.03192804008722305, + 0.02158314362168312, + 0.032334670424461365, + 0.033241402357816696, + 0.022273089736700058, + 0.03974615037441254, + 0.05446565896272659, + 0.05551709234714508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.009779755026102066 + }, + { + "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 static wide-angle view of a young girl standing on a concrete pedestrian bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl have short light brown hair and wear a white sailor-style shirt with a dark pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large full moon visible rising above the horizon, casting the girl as a silhouette with silver light rims on her hair and shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a reflection of the moon shimmering on the dark water of a river below the bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are soft, warm-toned city lights visible in the distance creating a bokeh effect?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634967729449272 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634967729449272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634967729449272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12474263459444046, + 0.03550342470407486, + 0.11792151629924774, + 0.01791619509458542, + 0.12812212109565735, + 0.09772385656833649, + 0.07657865434885025, + 0.03933236747980118, + 0.11405233293771744, + 0.2481069266796112 + ], + "credit_map": [ + 0.013762496411800385, + 0.010082405991852283, + 0.018635032698512077, + 0.009765149094164371, + 0.01700541004538536, + 0.015928547829389572, + 0.012756994925439358, + 0.026806587353348732, + 0.005613317713141441, + 0.0030033630318939686, + 0.0037994885351508856, + 0.0029936074279248714, + 0.003764863358810544, + 0.006504659075289965, + 0.003163072979077697, + 0.006661050952970982, + 0.010071143507957458, + 0.008569678291678429, + 0.014965186826884747, + 0.011125018820166588, + 0.009288894943892956, + 0.014359349384903908, + 0.022011052817106247, + 0.027531176805496216, + 0.002599160885438323, + 0.002214146312326193, + 0.001734066754579544, + 0.0018037221161648631, + 0.002258897293359041, + 0.002197506371885538, + 0.0023981467820703983, + 0.0027105470653623343, + 0.016737638041377068, + 0.010764209553599358, + 0.013245850801467896, + 0.01588711328804493, + 0.01210958044975996, + 0.015462538227438927, + 0.01624979078769684, + 0.027665387839078903, + 0.014820368029177189, + 0.01745041459798813, + 0.010095169767737389, + 0.007487011607736349, + 0.012438164092600346, + 0.01208567712455988, + 0.008560127578675747, + 0.014786913990974426, + 0.010988359339535236, + 0.007547255605459213, + 0.01231662929058075, + 0.006531304679811001, + 0.005313817411661148, + 0.009953519329428673, + 0.008039243519306183, + 0.015888508409261703, + 0.005486981011927128, + 0.0046586827374994755, + 0.004134411457926035, + 0.003264806931838393, + 0.004050551448017359, + 0.0058424463495612144, + 0.0067381346598267555, + 0.005156346596777439, + 0.0126600731164217, + 0.009410524740815163, + 0.01285783015191555, + 0.009882028214633465, + 0.00906743761152029, + 0.015960900112986565, + 0.017942683771252632, + 0.02627084031701088, + 0.027954187244176865, + 0.0172391626983881, + 0.02571669965982437, + 0.02646750770509243, + 0.01829431764781475, + 0.03387176990509033, + 0.046903785318136215, + 0.05165946111083031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.008634967729449272 + } + ], + "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/0000017/group_4/index_tensor([5900, 5900])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_tensor([5900, 5900])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4b338aaaa09b9daff96e0a70cc78f0dfe844e8de --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_tensor([5900, 5900])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle medium shot of a closely following vehicle then rapidly pulls back to reveal a dark green SUV, splattered with thick mud, jolting violently along a bumpy dirt mountain road. The massive all-terrain tires struggle with the water-filled ruts, sending mud and debris flying high behind them. The vehicle heaves and pitches violently with the terrain, straining the suspension to its limits. Under a gloomy sky, diffuse light accentuates the sticky texture of the wet mud on the paint. As the camera rapidly pulls back, the tiny SUV is suddenly placed within a vast and precipitous panorama. It navigates alone on a narrow path winding up a steep mountainside, surrounded by a continuous series of sharp, gray peaks, some of which are hidden in low-hanging clouds. The entire scene exudes a sense of grandeur and loneliness, a sense of conquering nature, with a cold palette and a sense of adventurous tension.", + "seed": 5846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_5900_seed_5846_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": -0.04000878045335412, + "vlm_logprob": -0.04000878045335412, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04000878045335412, + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10982140898704529 + ], + "opposite_logprob_tokens": [ + -2.859821319580078 + ], + "logprob_reward": -0.10982140898704529, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.10982140898704529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05575185641646385, + 0.03436046466231346, + 0.05463161692023277, + 0.1012774407863617, + 0.026388516649603844, + 0.021726172417402267, + 0.03342004492878914, + 0.534163236618042, + 0.09813667088747025, + 0.04014401510357857 + ], + "credit_map": [ + 0.007344134617596865, + 0.0068466332741081715, + 0.006331604439765215, + 0.007599606178700924, + 0.008586438372731209, + 0.00675415713340044, + 0.007234475575387478, + 0.005054804030805826, + 0.0036760831717401743, + 0.005076981149613857, + 0.0049072494730353355, + 0.0038645246531814337, + 0.0055380892008543015, + 0.0029441523365676403, + 0.003861663630232215, + 0.004491723142564297, + 0.007551208604127169, + 0.009076670743525028, + 0.005090615246444941, + 0.0037321492563933134, + 0.010703060775995255, + 0.005396763328462839, + 0.005763403605669737, + 0.00731774652376771, + 0.006972959265112877, + 0.0032545567955821753, + 0.006848517805337906, + 0.0027800637762993574, + 0.006031970959156752, + 0.03722425922751427, + 0.005415181629359722, + 0.03274993225932121, + 0.005037698429077864, + 0.002782627008855343, + 0.002490651560947299, + 0.0034770299680531025, + 0.002268221229314804, + 0.0033765765838325024, + 0.0034750814083963633, + 0.003480628365650773, + 0.004024828784167767, + 0.003550074528902769, + 0.0020660420414060354, + 0.002440494718030095, + 0.0020547860767692327, + 0.0021715317852795124, + 0.0028667105361819267, + 0.0025517032481729984, + 0.0033354852348566055, + 0.004500159528106451, + 0.004006751347333193, + 0.005394762847572565, + 0.00311684631742537, + 0.0049999775364995, + 0.005005957558751106, + 0.003060105489566922, + 0.06634046882390976, + 0.15451842546463013, + 0.040592219680547714, + 0.051431089639663696, + 0.07188887894153595, + 0.055803995579481125, + 0.038476500660181046, + 0.05511166527867317, + 0.012945313937962055, + 0.017732976004481316, + 0.007847440429031849, + 0.010968013666570187, + 0.01275181956589222, + 0.015988577157258987, + 0.008944849483668804, + 0.010957679711282253, + 0.0035303044132888317, + 0.005377907771617174, + 0.007360775955021381, + 0.0042803059332072735, + 0.00374290207400918, + 0.0054314653389155865, + 0.0053925346583127975, + 0.0050278170965611935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.10982140898704529 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310541972517967 + ], + "opposite_logprob_tokens": [ + -3.168105363845825 + ], + "logprob_reward": -0.04310541972517967, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310541972517967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06881178170442581, + 0.04878561198711395, + 0.04686286300420761, + 0.04971342533826828, + 0.015867237001657486, + 0.022453172132372856, + 0.06655512005090714, + 0.5302156805992126, + 0.10823029279708862, + 0.04250478744506836 + ], + "credit_map": [ + 0.007439295295625925, + 0.009511029347777367, + 0.012008942663669586, + 0.008450455963611603, + 0.009428121149539948, + 0.006394449155777693, + 0.008576340042054653, + 0.007003146223723888, + 0.004906453657895327, + 0.005245220381766558, + 0.007970582693815231, + 0.0066072773188352585, + 0.00775131955742836, + 0.005135567393153906, + 0.005244549363851547, + 0.005924637895077467, + 0.0064136856235563755, + 0.007587186526507139, + 0.005229447036981583, + 0.003715924918651581, + 0.007327083498239517, + 0.0062964968383312225, + 0.004638529848307371, + 0.005654509179294109, + 0.004863733891397715, + 0.0026272886898368597, + 0.0033852807246148586, + 0.002884156070649624, + 0.004301854409277439, + 0.015314489603042603, + 0.003655355889350176, + 0.012681269086897373, + 0.0035049470607191324, + 0.001994831021875143, + 0.001444221823476255, + 0.0019182206597179174, + 0.001816950854845345, + 0.002050535986199975, + 0.0016325750621035695, + 0.0015049534849822521, + 0.003234623232856393, + 0.002460148185491562, + 0.0018923137104138732, + 0.0025168966967612505, + 0.003248209599405527, + 0.0029372319113463163, + 0.0037029385566711426, + 0.0024608103558421135, + 0.008436930365860462, + 0.008618470281362534, + 0.005883020348846912, + 0.010433577001094818, + 0.006540218833833933, + 0.011389867402613163, + 0.008181709796190262, + 0.0070713297463953495, + 0.0665900781750679, + 0.14257512986660004, + 0.06466221809387207, + 0.0470317043364048, + 0.06675635278224945, + 0.051073137670755386, + 0.03485429286956787, + 0.05667278170585632, + 0.02016020193696022, + 0.018562454730272293, + 0.00972759909927845, + 0.01006285473704338, + 0.012410219758749008, + 0.012434713542461395, + 0.010514291003346443, + 0.01435795333236456, + 0.0029527221340686083, + 0.004603681154549122, + 0.0055166855454444885, + 0.007394286338239908, + 0.006426173262298107, + 0.0041719102300703526, + 0.007172980811446905, + 0.004266346339136362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.04310541972517967 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590949803590775 + ], + "opposite_logprob_tokens": [ + -3.8955910205841064 + ], + "logprob_reward": -0.020590949803590775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590949803590775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024705350399017334, + 0.033247366547584534, + 0.020996300503611565, + 0.021399499848484993, + 0.018058011308312416, + 0.019463589414954185, + 0.06165419518947601, + 0.6467961072921753, + 0.12275360524654388, + 0.030925964936614037 + ], + "credit_map": [ + 0.0025843216571956873, + 0.002789073158055544, + 0.00395432161167264, + 0.002598782069981098, + 0.0038345663342624903, + 0.003065643599256873, + 0.003093540435656905, + 0.002785101532936096, + 0.0032605885062366724, + 0.0032445068936794996, + 0.006437714211642742, + 0.004232858307659626, + 0.005079431459307671, + 0.00397353945299983, + 0.0032033822499215603, + 0.0038153475616127253, + 0.002845353912562132, + 0.0033339199144393206, + 0.002153912093490362, + 0.00173846329562366, + 0.00437231594696641, + 0.0020659733563661575, + 0.002278489526361227, + 0.002207872923463583, + 0.003803960280492902, + 0.0022270516492426395, + 0.0017103737918660045, + 0.0014970145421102643, + 0.004825438838452101, + 0.002845891984179616, + 0.002821665722876787, + 0.0016681038541719317, + 0.003962123300880194, + 0.002224718453362584, + 0.0019521057838574052, + 0.001275347312912345, + 0.0014102786080911756, + 0.002003585686907172, + 0.0015963665209710598, + 0.0036334849428385496, + 0.003414066741243005, + 0.001829013112001121, + 0.0017782286740839481, + 0.0027445813175290823, + 0.002405846491456032, + 0.002480434952303767, + 0.0032071401365101337, + 0.0016042797360569239, + 0.00648636557161808, + 0.014882911927998066, + 0.0045637390576303005, + 0.0066046747379004955, + 0.0068994369357824326, + 0.008015810512006283, + 0.005710143595933914, + 0.008491113781929016, + 0.07808561623096466, + 0.1740080714225769, + 0.07998556643724442, + 0.05737245827913284, + 0.08147979527711868, + 0.06412167847156525, + 0.044662803411483765, + 0.06708011031150818, + 0.011592336930334568, + 0.0214095376431942, + 0.00981925893574953, + 0.01789344660937786, + 0.016392171382904053, + 0.020963819697499275, + 0.012173603288829327, + 0.012509431689977646, + 0.002079467289149761, + 0.002662641927599907, + 0.0035792405251413584, + 0.003496853867545724, + 0.005571198649704456, + 0.005419496912509203, + 0.004420664627104998, + 0.003696402069181204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.020590949803590775 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10037945210933685, + 0.04835944622755051, + 0.04748049005866051, + 0.23506926000118256, + 0.07252131402492523, + 0.09708697348833084, + 0.07182931154966354, + 0.16154871881008148, + 0.11752737313508987, + 0.04819773882627487 + ], + "credit_map": [ + 0.01257769763469696, + 0.016429658979177475, + 0.013772708363831043, + 0.008655295707285404, + 0.015204224735498428, + 0.012564172968268394, + 0.01087086834013462, + 0.0103048300370574, + 0.004966783802956343, + 0.007879653014242649, + 0.0069309924729168415, + 0.00566023588180542, + 0.007300487719476223, + 0.004412660840898752, + 0.0060639167204499245, + 0.005144715309143066, + 0.005261266604065895, + 0.006874921731650829, + 0.0045655472204089165, + 0.0038865269161760807, + 0.009324035607278347, + 0.006074517033994198, + 0.005832660477608442, + 0.00566101586446166, + 0.016231006011366844, + 0.0051689064130187035, + 0.01650487445294857, + 0.004993991926312447, + 0.013064028695225716, + 0.0867285430431366, + 0.014433866366744041, + 0.07794404774904251, + 0.010117249563336372, + 0.011413727886974812, + 0.008328049443662167, + 0.009076918475329876, + 0.00714986864477396, + 0.01053400244563818, + 0.008442135527729988, + 0.007459360174834728, + 0.017830360680818558, + 0.02007460407912731, + 0.009554148651659489, + 0.0076723163947463036, + 0.010614024475216866, + 0.011272687464952469, + 0.007788066752254963, + 0.012280761264264584, + 0.007041952572762966, + 0.014817710034549236, + 0.006963123567402363, + 0.008936944417655468, + 0.00808274932205677, + 0.008614525198936462, + 0.008930494077503681, + 0.008441810496151447, + 0.011170064099133015, + 0.01876986399292946, + 0.04399440065026283, + 0.01483826246112585, + 0.01105239987373352, + 0.018639517948031425, + 0.013215307146310806, + 0.029868904501199722, + 0.016054632142186165, + 0.02093871682882309, + 0.008122206665575504, + 0.014147755689918995, + 0.016598066315054893, + 0.017808686941862106, + 0.011955004185438156, + 0.01190231367945671, + 0.004086450207978487, + 0.009097350761294365, + 0.010844207368791103, + 0.003432614728808403, + 0.004267296753823757, + 0.006645982153713703, + 0.003466229187324643, + 0.006357607431709766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.02057565003633499 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950473714619875 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950473714619875, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950473714619875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062192969024181366, + 0.014715724624693394, + 0.029239974915981293, + 0.5212811827659607, + 0.02806048095226288, + 0.02346303127706051, + 0.0373104028403759, + 0.1756173074245453, + 0.08365541696548462, + 0.024463502690196037 + ], + "credit_map": [ + 0.010024520568549633, + 0.0069318474270403385, + 0.009428085759282112, + 0.004765181802213192, + 0.00822561327368021, + 0.008669033646583557, + 0.005984173621982336, + 0.008164500817656517, + 0.0015154264401644468, + 0.0017968275351449847, + 0.002478857757523656, + 0.001882354961708188, + 0.0014074587961658835, + 0.0020963253919035196, + 0.0018555686110630631, + 0.0016829033847898245, + 0.004071858245879412, + 0.0030448497273027897, + 0.0035725408233702183, + 0.0021007617469877005, + 0.003943812567740679, + 0.005334741435945034, + 0.0031474584247916937, + 0.004023948684334755, + 0.04091295972466469, + 0.011681991629302502, + 0.03675651550292969, + 0.009559743106365204, + 0.028105255216360092, + 0.19835633039474487, + 0.02507067658007145, + 0.17083770036697388, + 0.006604866124689579, + 0.003217540681362152, + 0.002526555210351944, + 0.002616677898913622, + 0.0018099792068824172, + 0.0037390547804534435, + 0.0020371819846332073, + 0.005508623085916042, + 0.0040381187573075294, + 0.0031645160634070635, + 0.0022092321887612343, + 0.003079737303778529, + 0.002494063461199403, + 0.002726515755057335, + 0.0034711898770183325, + 0.0022796557750552893, + 0.004268964286893606, + 0.0055719963274896145, + 0.003823918290436268, + 0.004366749431937933, + 0.0032519365195184946, + 0.004271300975233316, + 0.005641923286020756, + 0.006113613490015268, + 0.02153932861983776, + 0.04413392022252083, + 0.023605888709425926, + 0.015938200056552887, + 0.021449822932481766, + 0.016135532408952713, + 0.013158026151359081, + 0.019656578078866005, + 0.012604099698364735, + 0.014437994919717312, + 0.006320230197161436, + 0.00915787648409605, + 0.011029495857656002, + 0.012406207621097565, + 0.007812869735062122, + 0.009886636398732662, + 0.002450196770951152, + 0.0030829072929918766, + 0.003831839421764016, + 0.002666913438588381, + 0.003185389330610633, + 0.003358895890414715, + 0.0033249491825699806, + 0.0025624071713536978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.005950473714619875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle medium shot of a closely following vehicle then rapidly pulls back to reveal a dark green SUV, splattered with thick mud, jolting violently along a bumpy dirt mountain road. The massive all-terrain tires struggle with the water-filled ruts, sending mud and debris flying high behind them. The vehicle heaves and pitches violently with the terrain, straining the suspension to its limits. Under a gloomy sky, diffuse light accentuates the sticky texture of the wet mud on the paint. As the camera rapidly pulls back, the tiny SUV is suddenly placed within a vast and precipitous panorama. It navigates alone on a narrow path winding up a steep mountainside, surrounded by a continuous series of sharp, gray peaks, some of which are hidden in low-hanging clouds. The entire scene exudes a sense of grandeur and loneliness, a sense of conquering nature, with a cold palette and a sense of adventurous tension.", + "seed": 5847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_5900_seed_5847_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": -0.027241590432822704, + "vlm_logprob": -0.027241590432822704, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.027241590432822704, + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08504777401685715 + ], + "opposite_logprob_tokens": [ + -2.835047721862793 + ], + "logprob_reward": -0.08504777401685715, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08504777401685715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1597844660282135, + 0.07203647494316101, + 0.05447075888514519, + 0.10540757328271866, + 0.14280207455158234, + 0.11240911483764648, + 0.1723717600107193, + 0.10365167260169983, + 0.025346174836158752, + 0.051719896495342255 + ], + "credit_map": [ + 0.015316125005483627, + 0.0172471571713686, + 0.01854032464325428, + 0.017975080758333206, + 0.018136614933609962, + 0.015262304805219173, + 0.04035380110144615, + 0.016953062266111374, + 0.00801040604710579, + 0.005991247016936541, + 0.011803331784904003, + 0.007242810446768999, + 0.005245236214250326, + 0.0057958802208304405, + 0.01639094389975071, + 0.011556623503565788, + 0.004292108118534088, + 0.005234261509031057, + 0.009428710676729679, + 0.009634329006075859, + 0.0072940620593726635, + 0.00379375321790576, + 0.007491142489016056, + 0.007302391342818737, + 0.006686566863209009, + 0.00839188788086176, + 0.01573736034333706, + 0.01640711911022663, + 0.02156202122569084, + 0.008119598031044006, + 0.01501440815627575, + 0.013488617725670338, + 0.008409634232521057, + 0.009391311556100845, + 0.016651300713419914, + 0.02230215258896351, + 0.02017756551504135, + 0.010359641164541245, + 0.024062344804406166, + 0.0314481221139431, + 0.005624162033200264, + 0.009188312105834484, + 0.017081283032894135, + 0.01610727235674858, + 0.015611507929861546, + 0.009406452998518944, + 0.02142971195280552, + 0.017960410565137863, + 0.010583914816379547, + 0.013486242853105068, + 0.025199539959430695, + 0.02857111394405365, + 0.02250041998922825, + 0.011828023940324783, + 0.032349102199077606, + 0.027853410691022873, + 0.00517208781093359, + 0.007169107440859079, + 0.012525857426226139, + 0.01286974735558033, + 0.010580415837466717, + 0.010100165382027626, + 0.025598276406526566, + 0.019636020064353943, + 0.002390013774856925, + 0.002017262391746044, + 0.0037955683656036854, + 0.003471346804872155, + 0.003640382084995508, + 0.003547790227457881, + 0.00399394566193223, + 0.002489865757524967, + 0.0057991608045995235, + 0.0042981491424143314, + 0.00952715240418911, + 0.0054392158053815365, + 0.007010793779045343, + 0.005756800062954426, + 0.00618370994925499, + 0.00770491361618042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.08504777401685715 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062469720840454 + ], + "opposite_logprob_tokens": [ + -3.8956246376037598 + ], + "logprob_reward": -0.02062469720840454, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062469720840454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10699941217899323, + 0.12353038787841797, + 0.0982571467757225, + 0.13719409704208374, + 0.10476900637149811, + 0.14873428642749786, + 0.09265435487031937, + 0.13233283162117004, + 0.040028732270002365, + 0.015499662607908249 + ], + "credit_map": [ + 0.01122367661446333, + 0.011474988423287868, + 0.015810782089829445, + 0.014862625859677792, + 0.008454789407551289, + 0.007200528867542744, + 0.0237735603004694, + 0.01419847086071968, + 0.008738048374652863, + 0.009228612296283245, + 0.02000678889453411, + 0.015188626013696194, + 0.009051989763975143, + 0.00833615381270647, + 0.03328416496515274, + 0.019696014001965523, + 0.007329920306801796, + 0.010021851398050785, + 0.016750380396842957, + 0.014753450639545918, + 0.015357904136180878, + 0.007741162553429604, + 0.014165245927870274, + 0.01213724073022604, + 0.007640442810952663, + 0.011423065327107906, + 0.020902549847960472, + 0.011905163526535034, + 0.03066261112689972, + 0.012537889182567596, + 0.027291754260659218, + 0.014830627478659153, + 0.005768629722297192, + 0.007707890123128891, + 0.013988413847982883, + 0.018483836203813553, + 0.017293090000748634, + 0.008502734825015068, + 0.014339813962578773, + 0.018684616312384605, + 0.0063667818903923035, + 0.012642564252018929, + 0.023120980709791183, + 0.020063325762748718, + 0.024686157703399658, + 0.010715794749557972, + 0.03007371351122856, + 0.02106497436761856, + 0.0052041965536773205, + 0.007449394557625055, + 0.015355348587036133, + 0.016648398712277412, + 0.010222917422652245, + 0.00836489163339138, + 0.015983011573553085, + 0.013426200486719608, + 0.00711093470454216, + 0.00957521889358759, + 0.01708563230931759, + 0.017811082303524017, + 0.013153312727808952, + 0.01037887018173933, + 0.029400460422039032, + 0.027817336842417717, + 0.0030365968123078346, + 0.0034692524932324886, + 0.005388451740145683, + 0.005311225075274706, + 0.006900270003825426, + 0.003843240672722459, + 0.008251414634287357, + 0.0038282866589725018, + 0.00158085604198277, + 0.0013503259979188442, + 0.0022696780506521463, + 0.002018902450799942, + 0.0015158505411818624, + 0.0015444675227627158, + 0.0025784396566450596, + 0.0026411437429487705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.02062469720840454 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1949906349182129, + 0.12988567352294922, + 0.1152179017663002, + 0.12397363781929016, + 0.09198584407567978, + 0.06286394596099854, + 0.08581879734992981, + 0.06666463613510132, + 0.07306531816720963, + 0.055533599108457565 + ], + "credit_map": [ + 0.024897240102291107, + 0.024296190589666367, + 0.03123563528060913, + 0.031986434012651443, + 0.015818284824490547, + 0.015296926721930504, + 0.030027827247977257, + 0.02143207937479019, + 0.01027059368789196, + 0.011122745461761951, + 0.019691625609993935, + 0.01427887286990881, + 0.013161624781787395, + 0.009759938344359398, + 0.029904987663030624, + 0.02169528789818287, + 0.00713265361264348, + 0.011592860333621502, + 0.0190742127597332, + 0.014034290798008442, + 0.022905882447957993, + 0.00978823285549879, + 0.018202252686023712, + 0.012487524189054966, + 0.006001048255711794, + 0.010812421329319477, + 0.01993536204099655, + 0.01745920069515705, + 0.021142657846212387, + 0.00892811082303524, + 0.024213487282395363, + 0.015481355600059032, + 0.005367680452764034, + 0.007450536359101534, + 0.01422070898115635, + 0.01593763567507267, + 0.008356732316315174, + 0.006629797630012035, + 0.0108704324811697, + 0.023152314126491547, + 0.006662100087851286, + 0.004764700774103403, + 0.006635772995650768, + 0.011017652228474617, + 0.010987208224833012, + 0.0037481363397091627, + 0.011732266284525394, + 0.007316115312278271, + 0.00548413023352623, + 0.006951683200895786, + 0.011918292380869389, + 0.011125284247100353, + 0.013456546701490879, + 0.0070436266250908375, + 0.018373757600784302, + 0.01146547682583332, + 0.0033113451208919287, + 0.004682626109570265, + 0.009783703833818436, + 0.011372247710824013, + 0.005710735451430082, + 0.005421879701316357, + 0.013182157650589943, + 0.01319994404911995, + 0.005037667229771614, + 0.005993678234517574, + 0.010089751332998276, + 0.010081959888339043, + 0.010542377829551697, + 0.008071620017290115, + 0.014644543640315533, + 0.008603721857070923, + 0.006046673748642206, + 0.005077725742012262, + 0.007286405190825462, + 0.006324402987957001, + 0.0070026712492108345, + 0.007706007454544306, + 0.007877454161643982, + 0.0082122553139925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014205172657966614 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522008776664734, + 0.07864485681056976, + 0.0631878674030304, + 0.11188014596700668, + 0.1332167536020279, + 0.04892345517873764, + 0.07960478961467743, + 0.21453185379505157, + 0.04242929443717003, + 0.07538005709648132 + ], + "credit_map": [ + 0.0201108381152153, + 0.017300773411989212, + 0.021879058331251144, + 0.020060360431671143, + 0.02320932224392891, + 0.013696352951228619, + 0.01588483899831772, + 0.02005934901535511, + 0.005690103862434626, + 0.006883515976369381, + 0.017183268442749977, + 0.009371778927743435, + 0.008211228996515274, + 0.004931253381073475, + 0.014326557517051697, + 0.012047160416841507, + 0.006004442926496267, + 0.00526990182697773, + 0.014032355509698391, + 0.012739942409098148, + 0.0068061621859669685, + 0.004671359900385141, + 0.006034131161868572, + 0.007629574742168188, + 0.01301318034529686, + 0.010274034924805164, + 0.01765858195722103, + 0.019399749115109444, + 0.016355032101273537, + 0.008377403020858765, + 0.015752427279949188, + 0.011049742810428143, + 0.008346379734575748, + 0.01076706126332283, + 0.020291725173592567, + 0.029645422473549843, + 0.012573400512337685, + 0.009046748280525208, + 0.012319209985435009, + 0.0302268099039793, + 0.0050963968969881535, + 0.004320893436670303, + 0.007586839143186808, + 0.009106653742492199, + 0.006280466448515654, + 0.0041793519631028175, + 0.005802324507385492, + 0.006550534628331661, + 0.005562698934227228, + 0.006887434981763363, + 0.011837582103908062, + 0.012822205200791359, + 0.010349762625992298, + 0.007730057463049889, + 0.014874428510665894, + 0.00954062957316637, + 0.014276533387601376, + 0.020238852128386497, + 0.03833150863647461, + 0.02958129160106182, + 0.03461109846830368, + 0.02532503940165043, + 0.02197367697954178, + 0.030193863436579704, + 0.0023315215948969126, + 0.003396144136786461, + 0.004993323236703873, + 0.004974743351340294, + 0.0063393609598279, + 0.0051809875294566154, + 0.010344015434384346, + 0.0048692054115235806, + 0.01027172151952982, + 0.0073613012209534645, + 0.012734721414744854, + 0.009268217720091343, + 0.0076325139962136745, + 0.01144389621913433, + 0.0079385656863451, + 0.008729123510420322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.011075802147388458 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254506133496761 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254506133496761, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254506133496761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18680095672607422, + 0.09326900541782379, + 0.10190683603286743, + 0.09574225544929504, + 0.0815509483218193, + 0.09005678445100784, + 0.0832105427980423, + 0.1015513613820076, + 0.1055363342165947, + 0.06037495657801628 + ], + "credit_map": [ + 0.023900168016552925, + 0.024615518748760223, + 0.03335944563150406, + 0.020918576046824455, + 0.024346066638827324, + 0.014152636751532555, + 0.028184989467263222, + 0.017323538661003113, + 0.007204925175756216, + 0.009625589475035667, + 0.016680486500263214, + 0.014280518516898155, + 0.010861092247068882, + 0.00810946524143219, + 0.010218104347586632, + 0.01628882624208927, + 0.00866701453924179, + 0.009600356221199036, + 0.016377964988350868, + 0.020451027899980545, + 0.015562223270535469, + 0.007669437676668167, + 0.011009505949914455, + 0.012569312937557697, + 0.005552137270569801, + 0.009052986279129982, + 0.012778260745108128, + 0.018526598811149597, + 0.0169968381524086, + 0.007238584570586681, + 0.015679333359003067, + 0.009917518123984337, + 0.0037110387347638607, + 0.008188058622181416, + 0.012522258795797825, + 0.021202897652983665, + 0.010982287116348743, + 0.007154769729822874, + 0.008489777334034443, + 0.009299861267209053, + 0.004768671467900276, + 0.006781197618693113, + 0.014077655039727688, + 0.019857076928019524, + 0.012478264980018139, + 0.006065181456506252, + 0.01248688530176878, + 0.01354185026139021, + 0.005576475989073515, + 0.007633414585143328, + 0.012015688233077526, + 0.011502813547849655, + 0.012223782017827034, + 0.007426333148032427, + 0.01644577458500862, + 0.01038625743240118, + 0.006359125953167677, + 0.007268807850778103, + 0.013851944357156754, + 0.015551467426121235, + 0.012129376642405987, + 0.008962084539234638, + 0.02220936492085457, + 0.015219195745885372, + 0.00602682214230299, + 0.008056269958615303, + 0.01257360354065895, + 0.01080863457173109, + 0.015572283416986465, + 0.01222657784819603, + 0.02741866186261177, + 0.012853485532104969, + 0.007659445516765118, + 0.005683882161974907, + 0.010524402372539043, + 0.006991925649344921, + 0.007080675568431616, + 0.006844505202025175, + 0.007633249741047621, + 0.007956868037581444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.005254506133496761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle medium shot of a closely following vehicle then rapidly pulls back to reveal a dark green SUV, splattered with thick mud, jolting violently along a bumpy dirt mountain road. The massive all-terrain tires struggle with the water-filled ruts, sending mud and debris flying high behind them. The vehicle heaves and pitches violently with the terrain, straining the suspension to its limits. Under a gloomy sky, diffuse light accentuates the sticky texture of the wet mud on the paint. As the camera rapidly pulls back, the tiny SUV is suddenly placed within a vast and precipitous panorama. It navigates alone on a narrow path winding up a steep mountainside, surrounded by a continuous series of sharp, gray peaks, some of which are hidden in low-hanging clouds. The entire scene exudes a sense of grandeur and loneliness, a sense of conquering nature, with a cold palette and a sense of adventurous tension.", + "seed": 5848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_5900_seed_5848_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": -0.0189219253603369, + "vlm_logprob": -0.0189219253603369, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0189219253603369, + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04142991080880165 + ], + "opposite_logprob_tokens": [ + -4.0414299964904785 + ], + "logprob_reward": -0.04142991080880165, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04142991080880165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08296097815036774, + 0.04775646701455116, + 0.23200102150440216, + 0.06116649881005287, + 0.29055893421173096, + 0.12146469205617905, + 0.08281050622463226, + 0.027959508821368217, + 0.027727389708161354, + 0.02559402957558632 + ], + "credit_map": [ + 0.008944501169025898, + 0.012278707697987556, + 0.013846494257450104, + 0.009573801420629025, + 0.009388171136379242, + 0.01007222943007946, + 0.007816607132554054, + 0.0110404621809721, + 0.005425478797405958, + 0.00435934541746974, + 0.006702372804284096, + 0.006661944091320038, + 0.0064282831735908985, + 0.0056298356503248215, + 0.006787560880184174, + 0.005761646665632725, + 0.01413397490978241, + 0.009141980670392513, + 0.010025848634541035, + 0.01834343746304512, + 0.017851712182164192, + 0.07361594587564468, + 0.011540728621184826, + 0.07734737545251846, + 0.006720899138599634, + 0.005875557195395231, + 0.006344282068312168, + 0.004178153350949287, + 0.011158287525177002, + 0.010043283924460411, + 0.008093927055597305, + 0.008752112276852131, + 0.08019015192985535, + 0.014863681048154831, + 0.029623333364725113, + 0.02414965070784092, + 0.017942851409316063, + 0.02377152256667614, + 0.02201201766729355, + 0.07800573110580444, + 0.01669599860906601, + 0.012681441381573677, + 0.02065611258149147, + 0.0127484742552042, + 0.009197314269840717, + 0.025468554347753525, + 0.012561536394059658, + 0.011455265805125237, + 0.009692122228443623, + 0.009173110127449036, + 0.013799536973237991, + 0.008379016071557999, + 0.00810985080897808, + 0.011892499402165413, + 0.013108239509165287, + 0.008656132034957409, + 0.0025074102450162172, + 0.0031326001044362783, + 0.0038643116131424904, + 0.0041717044077813625, + 0.004480757750570774, + 0.0029008027631789446, + 0.004158603027462959, + 0.0027433172799646854, + 0.002291611861437559, + 0.003154696198180318, + 0.005240125581622124, + 0.005324851255863905, + 0.004251900594681501, + 0.0026967322919517756, + 0.0029834348242729902, + 0.001784037216566503, + 0.0021670933347195387, + 0.0026962091214954853, + 0.003947589546442032, + 0.003481462597846985, + 0.0036851672921329737, + 0.002729226602241397, + 0.003829324385151267, + 0.0030579562298953533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.04142991080880165 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652373053133488 + ], + "opposite_logprob_tokens": [ + -4.882652282714844 + ], + "logprob_reward": -0.007652373053133488, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652373053133488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08296947181224823, + 0.05582587793469429, + 0.16799227893352509, + 0.05103970319032669, + 0.23814189434051514, + 0.042905211448669434, + 0.1917397528886795, + 0.05661680921912193, + 0.06354853510856628, + 0.049220480024814606 + ], + "credit_map": [ + 0.007615553680807352, + 0.010658495128154755, + 0.017090754583477974, + 0.011553431861102581, + 0.0070898099802434444, + 0.007197718136012554, + 0.00987255945801735, + 0.011891153641045094, + 0.006237885449081659, + 0.007515306584537029, + 0.006613622419536114, + 0.010307203978300095, + 0.004551399499177933, + 0.005264893639832735, + 0.008151404559612274, + 0.007184162270277739, + 0.010311074554920197, + 0.007956811226904392, + 0.005901151802390814, + 0.015165304765105247, + 0.009134281426668167, + 0.0530303455889225, + 0.008739704266190529, + 0.05775360390543938, + 0.005600020755082369, + 0.00556491082534194, + 0.006186694372445345, + 0.004129465669393539, + 0.010358021594583988, + 0.006379891652613878, + 0.005665406119078398, + 0.007155291736125946, + 0.07067669183015823, + 0.008390567265450954, + 0.021815305575728416, + 0.013672646135091782, + 0.01299143023788929, + 0.019852448254823685, + 0.017822956666350365, + 0.07291984558105469, + 0.005883986596018076, + 0.004225656855851412, + 0.00866937916725874, + 0.004770844243466854, + 0.00485738692805171, + 0.004654621705412865, + 0.004159980919212103, + 0.005683355499058962, + 0.015936831012368202, + 0.017807913944125175, + 0.03017989546060562, + 0.027057671919465065, + 0.016610557213425636, + 0.018334051594138145, + 0.04485049471259117, + 0.02096232958137989, + 0.005580938886851072, + 0.006651849485933781, + 0.008756740018725395, + 0.007203607354313135, + 0.009130865335464478, + 0.004599214065819979, + 0.00825873576104641, + 0.0064348564483225346, + 0.003787474473938346, + 0.006341463420540094, + 0.011393564753234386, + 0.00936655979603529, + 0.009025806561112404, + 0.00533259054645896, + 0.011457971297204494, + 0.006843104958534241, + 0.004834681283682585, + 0.005992763210088015, + 0.006926396396011114, + 0.0053368667140603065, + 0.01058132667094469, + 0.005714695900678635, + 0.005978066474199295, + 0.003855682909488678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.007652373053133488 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802139848470688 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802139848470688, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802139848470688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045926354825496674, + 0.040954891592264175, + 0.33568868041038513, + 0.07729669660329819, + 0.13501891493797302, + 0.16116929054260254, + 0.07047443836927414, + 0.041746508330106735, + 0.0469987653195858, + 0.044725485146045685 + ], + "credit_map": [ + 0.005334837827831507, + 0.006014801096171141, + 0.009317790158092976, + 0.006640680134296417, + 0.004009339027106762, + 0.003917303867638111, + 0.005067944526672363, + 0.005623657256364822, + 0.004025625064969063, + 0.005047407001256943, + 0.0056795091368258, + 0.005998651031404734, + 0.0041274623945355415, + 0.0037640014197677374, + 0.007072212174534798, + 0.0052400194108486176, + 0.021563347429037094, + 0.015013369731605053, + 0.013459926471114159, + 0.03025864250957966, + 0.022139739245176315, + 0.09978676587343216, + 0.021662604063749313, + 0.11180426925420761, + 0.009224476292729378, + 0.005460639949887991, + 0.008045013062655926, + 0.0058787912130355835, + 0.010015467181801796, + 0.016612770035862923, + 0.010062814690172672, + 0.01199671346694231, + 0.032055359333753586, + 0.009087078273296356, + 0.017047856003046036, + 0.012761958874762058, + 0.009464251808822155, + 0.011085419915616512, + 0.011797633953392506, + 0.03171934559941292, + 0.020068705081939697, + 0.014864151366055012, + 0.03276620805263519, + 0.018838675692677498, + 0.012543155811727047, + 0.030238160863518715, + 0.016390148550271988, + 0.015460069291293621, + 0.008434975519776344, + 0.0079755038022995, + 0.010606731288135052, + 0.007305373437702656, + 0.008882613852620125, + 0.010365165770053864, + 0.00990817416459322, + 0.006995890289545059, + 0.0037114776205271482, + 0.004766186233609915, + 0.005859124008566141, + 0.005973798222839832, + 0.007079442031681538, + 0.0054349107667803764, + 0.00544202933087945, + 0.003479535458609462, + 0.003026549005880952, + 0.0045209163799881935, + 0.00840985681861639, + 0.009662373922765255, + 0.0048298584297299385, + 0.0037079458124935627, + 0.007812055759131908, + 0.005029202904552221, + 0.004442261531949043, + 0.004481601063162088, + 0.00644815806299448, + 0.005554209463298321, + 0.008647888898849487, + 0.005161821376532316, + 0.005797712132334709, + 0.004191828891634941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.029802139848470688 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533823028206825 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533823028206825, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533823028206825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09985534101724625, + 0.06808963418006897, + 0.1361495405435562, + 0.056709952652454376, + 0.14025844633579254, + 0.06286327540874481, + 0.1297311782836914, + 0.04376979544758797, + 0.07852791994810104, + 0.1840449422597885 + ], + "credit_map": [ + 0.007204818073660135, + 0.016440991312265396, + 0.013775639235973358, + 0.009662301279604435, + 0.012501135468482971, + 0.015439856797456741, + 0.009566469117999077, + 0.015264125540852547, + 0.005967112723737955, + 0.00709989108145237, + 0.00720813125371933, + 0.008973740972578526, + 0.008759506978094578, + 0.010028384625911713, + 0.009984539821743965, + 0.010068327188491821, + 0.013088702224195004, + 0.01210635993629694, + 0.009874559007585049, + 0.017094241455197334, + 0.013011880218982697, + 0.026612499728798866, + 0.01662975363433361, + 0.027731547132134438, + 0.005910441279411316, + 0.0036991399247199297, + 0.008677146397531033, + 0.00333409127779305, + 0.007464210502803326, + 0.00984483677893877, + 0.009146453812718391, + 0.008633633144199848, + 0.03610045090317726, + 0.008183126337826252, + 0.01545088179409504, + 0.012817535549402237, + 0.009116084314882755, + 0.01124552357941866, + 0.010844468139111996, + 0.03650037571787834, + 0.007490853779017925, + 0.005148840602487326, + 0.015853408724069595, + 0.008698306046426296, + 0.006867128424346447, + 0.006449040956795216, + 0.005427889060229063, + 0.006927805952727795, + 0.015586219727993011, + 0.01294617261737585, + 0.025021234527230263, + 0.012211030349135399, + 0.016737570986151695, + 0.022952377796173096, + 0.012572681531310081, + 0.01170390099287033, + 0.004338471684604883, + 0.005030556581914425, + 0.00686663668602705, + 0.006301280576735735, + 0.006754360161721706, + 0.005704635754227638, + 0.004651512484997511, + 0.004122345708310604, + 0.005829370580613613, + 0.008989064954221249, + 0.011631780304014683, + 0.01046824175864458, + 0.012502728961408138, + 0.012102888897061348, + 0.010727663524448872, + 0.006276178639382124, + 0.013861807994544506, + 0.021513430401682854, + 0.02176135778427124, + 0.013525791466236115, + 0.035106364637613297, + 0.02407699264585972, + 0.030361348763108253, + 0.02383786253631115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.012533823028206825 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191380063071847 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.003191380063071847, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191380063071847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10371799767017365, + 0.046708427369594574, + 0.21673642098903656, + 0.04699913412332535, + 0.06260567158460617, + 0.08910023421049118, + 0.07369011640548706, + 0.09274937957525253, + 0.09306973218917847, + 0.17462283372879028 + ], + "credit_map": [ + 0.011970839463174343, + 0.01171630434691906, + 0.01932470314204693, + 0.013315018266439438, + 0.009337959811091423, + 0.013606312684714794, + 0.010298050940036774, + 0.014148821122944355, + 0.004606322385370731, + 0.006326695438474417, + 0.006330349948257208, + 0.0073407432064414024, + 0.004335243720561266, + 0.0056663877330720425, + 0.007072510663419962, + 0.005030177533626556, + 0.014770517125725746, + 0.015714004635810852, + 0.012091347016394138, + 0.023721760138869286, + 0.01842889003455639, + 0.05338697135448456, + 0.02011188678443432, + 0.05851105973124504, + 0.005949828773736954, + 0.0054124630987644196, + 0.006606715265661478, + 0.0037063267081975937, + 0.00691821426153183, + 0.0059700096026062965, + 0.006339609157294035, + 0.006095972377806902, + 0.008228056132793427, + 0.008193385787308216, + 0.007846180349588394, + 0.007861541584134102, + 0.009650634601712227, + 0.007303628139197826, + 0.0075960466638207436, + 0.005926202982664108, + 0.010580508038401604, + 0.011262059211730957, + 0.014312214218080044, + 0.009339207783341408, + 0.0067076487466692924, + 0.014993256889283657, + 0.014333579689264297, + 0.007571765687316656, + 0.00599058298394084, + 0.006115504540503025, + 0.009904569014906883, + 0.011192543432116508, + 0.008302348665893078, + 0.008932787925004959, + 0.013892097398638725, + 0.009359683841466904, + 0.008028711192309856, + 0.010205634869635105, + 0.012290195561945438, + 0.011567560955882072, + 0.015287937596440315, + 0.00817850511521101, + 0.01558403018862009, + 0.011606813408434391, + 0.007575554773211479, + 0.0124589828774333, + 0.014842547476291656, + 0.010787748731672764, + 0.018582206219434738, + 0.008857584558427334, + 0.013900446705520153, + 0.006064676679670811, + 0.015601897612214088, + 0.02516314387321472, + 0.016890279948711395, + 0.012477213516831398, + 0.04397621378302574, + 0.01627655141055584, + 0.02880779281258583, + 0.01542975939810276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.003191380063071847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle medium shot of a closely following vehicle then rapidly pulls back to reveal a dark green SUV, splattered with thick mud, jolting violently along a bumpy dirt mountain road. The massive all-terrain tires struggle with the water-filled ruts, sending mud and debris flying high behind them. The vehicle heaves and pitches violently with the terrain, straining the suspension to its limits. Under a gloomy sky, diffuse light accentuates the sticky texture of the wet mud on the paint. As the camera rapidly pulls back, the tiny SUV is suddenly placed within a vast and precipitous panorama. It navigates alone on a narrow path winding up a steep mountainside, surrounded by a continuous series of sharp, gray peaks, some of which are hidden in low-hanging clouds. The entire scene exudes a sense of grandeur and loneliness, a sense of conquering nature, with a cold palette and a sense of adventurous tension.", + "seed": 5849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_4/index_5900_seed_5849_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": -0.02220816193148494, + "vlm_logprob": -0.02220816193148494, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.02220816193148494, + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04148983955383301 + ], + "opposite_logprob_tokens": [ + -4.041489601135254 + ], + "logprob_reward": -0.04148983955383301, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.04148983955383301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.127378910779953, + 0.06483129411935806, + 0.0959479883313179, + 0.13888485729694366, + 0.08334636688232422, + 0.04216446727514267, + 0.053129907697439194, + 0.3146507143974304, + 0.03789366036653519, + 0.041771821677684784 + ], + "credit_map": [ + 0.005919217597693205, + 0.0071744308806955814, + 0.04153779149055481, + 0.01039294246584177, + 0.01038411259651184, + 0.016653796657919884, + 0.01526104100048542, + 0.020055586472153664, + 0.006903194356709719, + 0.007047278806567192, + 0.009038416668772697, + 0.011019444093108177, + 0.005523037165403366, + 0.007573997601866722, + 0.008448821492493153, + 0.009277108125388622, + 0.009127862751483917, + 0.007682710886001587, + 0.010672200471162796, + 0.014966144226491451, + 0.014550713822245598, + 0.010403131134808064, + 0.013790661469101906, + 0.014754571951925755, + 0.015340982936322689, + 0.01537648681551218, + 0.011258918792009354, + 0.016260448843240738, + 0.0218135267496109, + 0.01481915544718504, + 0.03137252479791641, + 0.012642818503081799, + 0.007929807528853416, + 0.011526395566761494, + 0.009340042248368263, + 0.007518978323787451, + 0.009111418388783932, + 0.010097438469529152, + 0.02079727314412594, + 0.007025018334388733, + 0.008097746409475803, + 0.005619035568088293, + 0.005159668158739805, + 0.0028879737947136164, + 0.003392693819478154, + 0.008177660405635834, + 0.004958885721862316, + 0.003870809217914939, + 0.004762015771120787, + 0.004446389619261026, + 0.009676207788288593, + 0.008058506064116955, + 0.004439488053321838, + 0.00671240221709013, + 0.007692053914070129, + 0.007342847529798746, + 0.011807313188910484, + 0.007535804063081741, + 0.1334618180990219, + 0.018833264708518982, + 0.004516593646258116, + 0.03892986848950386, + 0.0070349774323403835, + 0.09253110736608505, + 0.0036880739498883486, + 0.002669382141903043, + 0.005270705558359623, + 0.005952662322670221, + 0.0023789010010659695, + 0.003308721585199237, + 0.006757940631359816, + 0.007867276668548584, + 0.0028470135293900967, + 0.0030158439185470343, + 0.006302960682660341, + 0.00582160335034132, + 0.0034130909480154514, + 0.003947177901864052, + 0.0081340866163373, + 0.00829005055129528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.04148983955383301 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00765769649296999 + ], + "opposite_logprob_tokens": [ + -4.882657527923584 + ], + "logprob_reward": -0.00765769649296999, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00765769649296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08043146133422852, + 0.10731986910104752, + 0.08812051266431808, + 0.13917772471904755, + 0.11071964353322983, + 0.24188978970050812, + 0.03177245706319809, + 0.12694697082042694, + 0.04558783397078514, + 0.02803378365933895 + ], + "credit_map": [ + 0.007014215923845768, + 0.00710352323949337, + 0.009848564863204956, + 0.016432536765933037, + 0.009705773554742336, + 0.008249539881944656, + 0.014661399647593498, + 0.007415909320116043, + 0.00968823954463005, + 0.01495162583887577, + 0.010503445751965046, + 0.013723875395953655, + 0.017764337360858917, + 0.016107946634292603, + 0.011447178199887276, + 0.013133218511939049, + 0.009428064338862896, + 0.009614778682589531, + 0.017428278923034668, + 0.011635194532573223, + 0.010886026546359062, + 0.011205676011741161, + 0.010250348597764969, + 0.007672149688005447, + 0.01317738275974989, + 0.017891108989715576, + 0.0109953498467803, + 0.018710345029830933, + 0.024696800857782364, + 0.017489956691861153, + 0.02354816347360611, + 0.01266860868781805, + 0.010182731784880161, + 0.011629943735897541, + 0.01513007190078497, + 0.012274001725018024, + 0.012780450284481049, + 0.012927217409014702, + 0.02569694258272648, + 0.010098285973072052, + 0.04522561654448509, + 0.0396215058863163, + 0.019339462742209435, + 0.014569067396223545, + 0.03319074213504791, + 0.041584983468055725, + 0.02473018877208233, + 0.023628227412700653, + 0.002655561314895749, + 0.0031812929082661867, + 0.004709692671895027, + 0.0062190028838813305, + 0.002400350756943226, + 0.003047951962798834, + 0.005688756238669157, + 0.003869847161695361, + 0.004725928418338299, + 0.00435367738828063, + 0.052239790558815, + 0.0075586517341434956, + 0.0026730725076049566, + 0.015702202916145325, + 0.005531178321689367, + 0.034162476658821106, + 0.004732026252895594, + 0.003716159611940384, + 0.00666958000510931, + 0.006480640731751919, + 0.0031799019780009985, + 0.004141001496464014, + 0.008444721810519695, + 0.008223803713917732, + 0.0019060558406636119, + 0.0026484245900064707, + 0.0039044299628585577, + 0.004081266466528177, + 0.0028221742250025272, + 0.0028517672326415777, + 0.005692875944077969, + 0.004126788582652807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.00765769649296999 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043065112084150314 + ], + "opposite_logprob_tokens": [ + -3.168065071105957 + ], + "logprob_reward": -0.043065112084150314, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043065112084150314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785508692264557, + 0.0398748479783535, + 0.10413151979446411, + 0.12505759298801422, + 0.06648784875869751, + 0.245208278298378, + 0.032330676913261414, + 0.1085132360458374, + 0.07638335973024368, + 0.02346181496977806 + ], + "credit_map": [ + 0.011808748356997967, + 0.007327808998525143, + 0.07132084667682648, + 0.017367307096719742, + 0.007871391251683235, + 0.020361606031656265, + 0.013177166692912579, + 0.02931600622832775, + 0.0048479498364031315, + 0.004470791667699814, + 0.004795231856405735, + 0.0037965690717101097, + 0.0049365670420229435, + 0.005057651083916426, + 0.006155401933938265, + 0.005814682692289352, + 0.009455962106585503, + 0.009117399342358112, + 0.02069396711885929, + 0.016224181279540062, + 0.01502953190356493, + 0.00855331402271986, + 0.013696707785129547, + 0.011360454373061657, + 0.013091734610497952, + 0.014534122310578823, + 0.0179936271160841, + 0.016000494360923767, + 0.018498988822102547, + 0.011924014426767826, + 0.020067282021045685, + 0.012947342358529568, + 0.007569571025669575, + 0.00840696133673191, + 0.010147289372980595, + 0.010744340717792511, + 0.006152275949716568, + 0.005905678030103445, + 0.01067984290421009, + 0.006881891284137964, + 0.06263411790132523, + 0.03448079153895378, + 0.018113266676664352, + 0.011330699548125267, + 0.027759701013565063, + 0.050618726760149, + 0.01886706054210663, + 0.02140391431748867, + 0.0036316358018666506, + 0.0029563980642706156, + 0.005169659852981567, + 0.005393600091338158, + 0.0027260982897132635, + 0.003304122481495142, + 0.005425353068858385, + 0.003723812522366643, + 0.003847678191959858, + 0.0023321742191910744, + 0.046721816062927246, + 0.0058027454651892185, + 0.0015496575506404042, + 0.013674543239176273, + 0.002795265521854162, + 0.03178935497999191, + 0.004153413698077202, + 0.003433381440117955, + 0.009741396643221378, + 0.02118195779621601, + 0.0030090503860265017, + 0.006769977509975433, + 0.013767520897090435, + 0.014326665550470352, + 0.0017245380440726876, + 0.0017868499271571636, + 0.0025503775104880333, + 0.003501392435282469, + 0.0013906698441132903, + 0.0022387399803847075, + 0.005547347944229841, + 0.0047218999825417995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.043065112084150314 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190010726451874 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190010726451874, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190010726451874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13688664138317108, + 0.037130746990442276, + 0.09979031980037689, + 0.04904155060648918, + 0.05226089060306549, + 0.11624778807163239, + 0.01991347409784794, + 0.3458535075187683, + 0.03914503753185272, + 0.10373011976480484 + ], + "credit_map": [ + 0.006400809623301029, + 0.0043612816371023655, + 0.05512464419007301, + 0.011861924082040787, + 0.0053033181466162205, + 0.022483251988887787, + 0.00874595157802105, + 0.022605441510677338, + 0.003953863866627216, + 0.004247670993208885, + 0.004078823607414961, + 0.005732374731451273, + 0.004003738984465599, + 0.004191039130091667, + 0.004819233901798725, + 0.006103998981416225, + 0.008195610716938972, + 0.007001310121268034, + 0.017418527975678444, + 0.018623504787683487, + 0.011791414581239223, + 0.010618701577186584, + 0.016204535961151123, + 0.009936694987118244, + 0.004365374334156513, + 0.005217670928686857, + 0.0048471796326339245, + 0.006973918993026018, + 0.006976086180657148, + 0.005096591543406248, + 0.009998536668717861, + 0.005566189531236887, + 0.006815181113779545, + 0.010750927031040192, + 0.008221650496125221, + 0.006352009251713753, + 0.0044564721174538136, + 0.0046502999030053616, + 0.006759996525943279, + 0.0042543490417301655, + 0.03819261118769646, + 0.01329043135046959, + 0.011276033706963062, + 0.004486671183258295, + 0.006114411633461714, + 0.027795586735010147, + 0.0054420423693954945, + 0.009649989195168018, + 0.0020751350093632936, + 0.0029728368390351534, + 0.003365248441696167, + 0.002702510915696621, + 0.0013884855434298515, + 0.0019207388395443559, + 0.0033506806939840317, + 0.0021378358360379934, + 0.012580646201968193, + 0.007825169712305069, + 0.14955611526966095, + 0.018595026805996895, + 0.0052239540964365005, + 0.043861620128154755, + 0.008262245915830135, + 0.0999487116932869, + 0.004015689250081778, + 0.003906978294253349, + 0.005642251577228308, + 0.006615940481424332, + 0.002430811058729887, + 0.003892490640282631, + 0.0055460380390286446, + 0.0070948353968560696, + 0.009235029108822346, + 0.010908869095146656, + 0.015007113106548786, + 0.022303635254502296, + 0.008245469070971012, + 0.010680698789656162, + 0.012856907211244106, + 0.014492388814687729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.014190010726451874 + }, + { + "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 initially a low-angle medium shot of a vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rapidly pull back to reveal a dark green SUV splattered with thick mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the SUV jolting violently on a bumpy dirt mountain road with water-filled ruts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast panorama of sharp gray peaks under a gloomy sky with low-hanging clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a cold color palette with diffuse light accentuating wet mud texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3990119397640228, + 0.03875770792365074, + 0.10710903257131577, + 0.08942914754152298, + 0.053500883281230927, + 0.09703335911035538, + 0.029141467064619064, + 0.08396615087985992, + 0.04195007309317589, + 0.06010032072663307 + ], + "credit_map": [ + 0.021181561052799225, + 0.01271622721105814, + 0.16694827377796173, + 0.03545414283871651, + 0.012444422580301762, + 0.05790882185101509, + 0.022888178005814552, + 0.06947031617164612, + 0.003650768892839551, + 0.0032161646522581577, + 0.004218962974846363, + 0.007902340032160282, + 0.00271484162658453, + 0.003556501120328903, + 0.006405800115317106, + 0.00709232734516263, + 0.012612075544893742, + 0.010170003399252892, + 0.010524140670895576, + 0.014842486940324306, + 0.017808372154831886, + 0.013967079110443592, + 0.017023969441652298, + 0.010160902515053749, + 0.007886219769716263, + 0.010716269724071026, + 0.009677161462605, + 0.008703600615262985, + 0.013159716501832008, + 0.009667603299021721, + 0.018789760768413544, + 0.010828814469277859, + 0.005952623672783375, + 0.007249584887176752, + 0.0057383268140256405, + 0.005010487046092749, + 0.006221890449523926, + 0.006665629800409079, + 0.009946583770215511, + 0.006715761031955481, + 0.03182816132903099, + 0.01061710063368082, + 0.00883403979241848, + 0.0034274477511644363, + 0.004734442103654146, + 0.02256348356604576, + 0.006683960556983948, + 0.008344721980392933, + 0.0031663274858146906, + 0.0036212617997080088, + 0.005681079812347889, + 0.003623130265623331, + 0.0024027579929679632, + 0.0027933414094150066, + 0.004522542003542185, + 0.0033310269936919212, + 0.003573955735191703, + 0.003114172024652362, + 0.03249732404947281, + 0.005540067330002785, + 0.0025043319910764694, + 0.010329077020287514, + 0.0050584860146045685, + 0.02134874276816845, + 0.002991721499711275, + 0.0034349653869867325, + 0.005181325599551201, + 0.00793544203042984, + 0.002200966002419591, + 0.0034011334646493196, + 0.006894279737025499, + 0.00991024263203144, + 0.005439084023237228, + 0.005968173500150442, + 0.008878801017999649, + 0.010351437143981457, + 0.004155084490776062, + 0.005606294609606266, + 0.007550010923296213, + 0.012151435017585754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.004638150800019503 + } + ], + "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/0000017/group_5/index_tensor([559, 559])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_tensor([559, 559])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3550ee88bd1c815920522c13702dcf020ad8b973 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_tensor([559, 559])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up begins a slow, mesmerizing push-in of a perfectly square block of pale golden butter placed in a dark, seasoned cast iron skillet. Under a warm overhead light, the butter's smooth, waxy surface begins to sweat, its sharp edges softening and collapsing as the heat takes hold. The solid form steadily dissolves from the bottom upward, giving way to an ever-expanding pool of shimmering, translucent liquid gold that catches the light with a brilliant sheen. Tiny, flaky bubbles form where the butter meets the hot pan, gently rising through the melting liquid before gently popping onto the surface. As the camera continues its deliberate forward movement, the last remnants of solid butter boil fully, transforming the scene into a gently bubbling, shimmering pool that ripples hypnotically, filling the frame with its rich, warm colors and creating a deeply satisfying and sensual visual experience.", + "seed": 19788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_559_seed_19788_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": -0.060952875204384326, + "vlm_logprob": -0.060952875204384326, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.060952875204384326, + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10482537001371384 + ], + "opposite_logprob_tokens": [ + -2.354825258255005 + ], + "logprob_reward": -0.10482537001371384, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10482537001371384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1456526517868042, + 0.07087185978889465, + 0.040205661207437515, + 0.11800600588321686, + 0.06355566531419754, + 0.05496867001056671, + 0.17370975017547607, + 0.1274225115776062, + 0.14188353717327118, + 0.06372366845607758 + ], + "credit_map": [ + 0.011646204628050327, + 0.009741347283124924, + 0.03547069802880287, + 0.026211276650428772, + 0.011429493315517902, + 0.014402097091078758, + 0.018658174201846123, + 0.01809336431324482, + 0.006239552982151508, + 0.005202359054237604, + 0.014463772997260094, + 0.009414696134626865, + 0.006504973396658897, + 0.008725267834961414, + 0.009352264925837517, + 0.010968973860144615, + 0.0034550968557596207, + 0.0037394969258457422, + 0.004152012523263693, + 0.004905233159661293, + 0.004529749974608421, + 0.003907782956957817, + 0.00527795497328043, + 0.010238333605229855, + 0.010101988911628723, + 0.009018650278449059, + 0.014261786825954914, + 0.021173667162656784, + 0.012688487768173218, + 0.01068736519664526, + 0.014420715160667896, + 0.025653347373008728, + 0.005214383359998465, + 0.00871105957776308, + 0.005304571241140366, + 0.0067670526914298534, + 0.008138865232467651, + 0.009423465467989445, + 0.00737452507019043, + 0.0126217445358634, + 0.0033265927340835333, + 0.011235058307647705, + 0.004867446143180132, + 0.005517903249710798, + 0.011210743337869644, + 0.00665750028565526, + 0.005759252700954676, + 0.0063941748812794685, + 0.02135893516242504, + 0.01869536191225052, + 0.020253848284482956, + 0.020510954782366753, + 0.021745605394244194, + 0.022750627249479294, + 0.02415786311030388, + 0.02423655427992344, + 0.013408883474767208, + 0.012852505780756474, + 0.017211996018886566, + 0.020833242684602737, + 0.014198639430105686, + 0.017027268186211586, + 0.015398417599499226, + 0.016491560265421867, + 0.012932557612657547, + 0.015970749780535698, + 0.033657047897577286, + 0.017554378136992455, + 0.013236882165074348, + 0.017070462927222252, + 0.013480979017913342, + 0.01798047497868538, + 0.006848410237580538, + 0.007653623819351196, + 0.010246165096759796, + 0.01055890042334795, + 0.005941036157310009, + 0.007696868386119604, + 0.007784685119986534, + 0.006993982009589672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.10482537001371384 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285357296466827 + ], + "opposite_logprob_tokens": [ + -2.0178534984588623 + ], + "logprob_reward": -0.14285357296466827, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285357296466827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0979066863656044, + 0.1634942591190338, + 0.07285498082637787, + 0.11350822448730469, + 0.08099759370088577, + 0.10134346783161163, + 0.08724875003099442, + 0.14427725970745087, + 0.049110714346170425, + 0.08925804495811462 + ], + "credit_map": [ + 0.008213651366531849, + 0.006806864868849516, + 0.02245546132326126, + 0.016346212476491928, + 0.009232839569449425, + 0.01138811931014061, + 0.009815191850066185, + 0.013648344203829765, + 0.015133467502892017, + 0.011408563703298569, + 0.02409300021827221, + 0.024508194997906685, + 0.011355563066899776, + 0.023081347346305847, + 0.02227674424648285, + 0.031637370586395264, + 0.008268702775239944, + 0.00524145970121026, + 0.009806717745959759, + 0.010491985827684402, + 0.005097813438624144, + 0.010283376090228558, + 0.007999862544238567, + 0.015665069222450256, + 0.010519125498831272, + 0.011322757229208946, + 0.009057234972715378, + 0.016163691878318787, + 0.010492396540939808, + 0.01603422686457634, + 0.01851499266922474, + 0.02140379510819912, + 0.00687249843031168, + 0.009784410707652569, + 0.008414044044911861, + 0.010205572471022606, + 0.009935988113284111, + 0.013324838131666183, + 0.009511558338999748, + 0.012948681600391865, + 0.008444427512586117, + 0.02247699536383152, + 0.007605902384966612, + 0.007613182533532381, + 0.022056180983781815, + 0.009853512048721313, + 0.010030487552285194, + 0.013262778520584106, + 0.006341495551168919, + 0.010357826948165894, + 0.007554835639894009, + 0.009365051984786987, + 0.012142903171479702, + 0.008333583362400532, + 0.010064572095870972, + 0.023088477551937103, + 0.016585713252425194, + 0.015735046938061714, + 0.01370550412684679, + 0.01888999156653881, + 0.017641158774495125, + 0.019899124279618263, + 0.01637103222310543, + 0.02544969692826271, + 0.004687354899942875, + 0.004857944790273905, + 0.006751163862645626, + 0.005774849094450474, + 0.005427320022135973, + 0.006365965586155653, + 0.006989605259150267, + 0.008256511762738228, + 0.012884690426290035, + 0.008599663153290749, + 0.012902265414595604, + 0.011937650851905346, + 0.008617403917014599, + 0.011434233747422695, + 0.01079258881509304, + 0.012089553289115429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.14285357296466827 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202288076281548 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.018202288076281548, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018202288076281548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22039969265460968, + 0.10179917514324188, + 0.05727227032184601, + 0.1382937729358673, + 0.12168344110250473, + 0.12289057672023773, + 0.08066999912261963, + 0.05319373309612274, + 0.058044880628585815, + 0.045752428472042084 + ], + "credit_map": [ + 0.021372390910983086, + 0.015143822878599167, + 0.0525885634124279, + 0.034329723566770554, + 0.019307216629385948, + 0.024881651625037193, + 0.02489188127219677, + 0.02788444235920906, + 0.009176324121654034, + 0.006537850480526686, + 0.021861104294657707, + 0.016799645498394966, + 0.006922426633536816, + 0.01207231730222702, + 0.01358058676123619, + 0.014848925173282623, + 0.005838955286890268, + 0.003545511746779084, + 0.007689403835684061, + 0.008203422650694847, + 0.004230321850627661, + 0.00513137923553586, + 0.007918024435639381, + 0.014715251512825489, + 0.011686718091368675, + 0.013114345259964466, + 0.016261380165815353, + 0.028535034507513046, + 0.013938616961240768, + 0.01587301306426525, + 0.016889285296201706, + 0.02199537865817547, + 0.010042399168014526, + 0.013784327544271946, + 0.010735295712947845, + 0.015486761927604675, + 0.015015398152172565, + 0.02349635399878025, + 0.01523481123149395, + 0.01788809336721897, + 0.0071515585295856, + 0.02364744059741497, + 0.010655266232788563, + 0.0115100909024477, + 0.02640991099178791, + 0.011716464534401894, + 0.01454014889895916, + 0.017259696498513222, + 0.009116467088460922, + 0.007200540974736214, + 0.007774315308779478, + 0.00954927783459425, + 0.010356479324400425, + 0.009120654314756393, + 0.011251719668507576, + 0.016300544142723083, + 0.0068322718143463135, + 0.007356183137744665, + 0.006962370127439499, + 0.006407686974853277, + 0.006170263513922691, + 0.0051915766671299934, + 0.005578941665589809, + 0.008694445714354515, + 0.007007008884102106, + 0.005457330495119095, + 0.011611813679337502, + 0.008407622575759888, + 0.00551427248865366, + 0.005894573405385017, + 0.007174152880907059, + 0.006978107150644064, + 0.006325547583401203, + 0.006733890622854233, + 0.006821985822170973, + 0.006212205626070499, + 0.005043384153395891, + 0.0037342426367104053, + 0.005044552031904459, + 0.005836621858179569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.018202288076281548 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541710399091244 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541710399091244, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541710399091244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14554469287395477, + 0.03390299901366234, + 0.10775259137153625, + 0.07273570448160172, + 0.16141431033611298, + 0.13979586958885193, + 0.06596235930919647, + 0.07173997908830643, + 0.12397342920303345, + 0.07717806100845337 + ], + "credit_map": [ + 0.012991813942790031, + 0.008598721586167812, + 0.028387589380145073, + 0.023721842095255852, + 0.011826439760625362, + 0.017189059406518936, + 0.020489756017923355, + 0.02233947068452835, + 0.0038720136508345604, + 0.0030426366720348597, + 0.004061145707964897, + 0.006451241672039032, + 0.003683972405269742, + 0.0038154644425958395, + 0.0036858199164271355, + 0.005290704313665628, + 0.011975129134953022, + 0.009105691686272621, + 0.015933604910969734, + 0.015683796256780624, + 0.009399225935339928, + 0.009271095506846905, + 0.012171720154583454, + 0.02421232871711254, + 0.0071778972633183, + 0.005213327240198851, + 0.00562767731025815, + 0.008678637444972992, + 0.006463466212153435, + 0.008181051351130009, + 0.011572792194783688, + 0.019820857793092728, + 0.011146973818540573, + 0.02067042700946331, + 0.01687021553516388, + 0.016468148678541183, + 0.021041082218289375, + 0.0235829409211874, + 0.01850324124097824, + 0.033131275326013565, + 0.009615219198167324, + 0.027859102934598923, + 0.01356471236795187, + 0.01388718094676733, + 0.02991282008588314, + 0.013101164251565933, + 0.01786411553621292, + 0.01399154681712389, + 0.006394514348357916, + 0.006996608804911375, + 0.008541090413928032, + 0.010102328844368458, + 0.009408978745341301, + 0.007380280643701553, + 0.007594444323331118, + 0.009544113650918007, + 0.009818557649850845, + 0.009702104143798351, + 0.009707053191959858, + 0.008730176836252213, + 0.008942710235714912, + 0.008227257989346981, + 0.0072965300641953945, + 0.009315586648881435, + 0.011842437088489532, + 0.013772733509540558, + 0.022700199857354164, + 0.016465842723846436, + 0.012038271874189377, + 0.016307448968291283, + 0.014569477178156376, + 0.016277018934488297, + 0.008358309976756573, + 0.008657015860080719, + 0.013730061240494251, + 0.011759157292544842, + 0.008071686141192913, + 0.00765089550986886, + 0.00875515304505825, + 0.010195787996053696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.012541710399091244 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341434568166733 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341434568166733, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341434568166733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17117591202259064, + 0.04244076833128929, + 0.08967563509941101, + 0.2625685930252075, + 0.09422672539949417, + 0.11716839671134949, + 0.08373900502920151, + 0.040794793516397476, + 0.07198166102170944, + 0.026228494942188263 + ], + "credit_map": [ + 0.012641947716474533, + 0.012764654122292995, + 0.03383631259202957, + 0.03390151634812355, + 0.016598550602793694, + 0.021220913156867027, + 0.021427500993013382, + 0.01878451742231846, + 0.0042868927121162415, + 0.0038472118321806192, + 0.00819570291787386, + 0.005918644834309816, + 0.004228520672768354, + 0.004725365899503231, + 0.004764943849295378, + 0.00647348677739501, + 0.007231057621538639, + 0.010421131737530231, + 0.009222217835485935, + 0.011898299679160118, + 0.010583738796412945, + 0.00997810997068882, + 0.012249440886080265, + 0.01809164136648178, + 0.02135222591459751, + 0.030196085572242737, + 0.026465389877557755, + 0.04144923761487007, + 0.02808762900531292, + 0.03062993288040161, + 0.03574392944574356, + 0.048644181340932846, + 0.008324175141751766, + 0.01480140071362257, + 0.009658856317400932, + 0.011321323923766613, + 0.017640886828303337, + 0.014630208723247051, + 0.0103227524086833, + 0.007527126930654049, + 0.013385345228016376, + 0.01048884354531765, + 0.012550726532936096, + 0.013246046379208565, + 0.013758868910372257, + 0.019676601514220238, + 0.015442739240825176, + 0.018619222566485405, + 0.007063472643494606, + 0.006457163020968437, + 0.007885736413300037, + 0.010644352063536644, + 0.009491154924035072, + 0.01280893012881279, + 0.011519677937030792, + 0.017868516966700554, + 0.004737740848213434, + 0.004932158626616001, + 0.003904029494151473, + 0.0055298213846981525, + 0.005482283886522055, + 0.004501250572502613, + 0.004226439632475376, + 0.007481069304049015, + 0.007409487850964069, + 0.007672775536775589, + 0.011273330077528954, + 0.01130822952836752, + 0.006701893173158169, + 0.008295905776321888, + 0.011379304341971874, + 0.007940735667943954, + 0.0027428369503468275, + 0.004215549677610397, + 0.003357416484504938, + 0.0037602323573082685, + 0.0032432996667921543, + 0.00264333956874907, + 0.003268745495006442, + 0.002997076138854027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.026341434568166733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, mesmerizing push-in of a perfectly square block of pale golden butter placed in a dark, seasoned cast iron skillet. Under a warm overhead light, the butter's smooth, waxy surface begins to sweat, its sharp edges softening and collapsing as the heat takes hold. The solid form steadily dissolves from the bottom upward, giving way to an ever-expanding pool of shimmering, translucent liquid gold that catches the light with a brilliant sheen. Tiny, flaky bubbles form where the butter meets the hot pan, gently rising through the melting liquid before gently popping onto the surface. As the camera continues its deliberate forward movement, the last remnants of solid butter boil fully, transforming the scene into a gently bubbling, shimmering pool that ripples hypnotically, filling the frame with its rich, warm colors and creating a deeply satisfying and sensual visual experience.", + "seed": 19789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_559_seed_19789_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": -0.01976667921990156, + "vlm_logprob": -0.01976667921990156, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.01976667921990156, + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012278917245566845 + ], + "opposite_logprob_tokens": [ + -4.762279033660889 + ], + "logprob_reward": -0.012278917245566845, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012278917245566845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10171445459127426, + 0.030364830046892166, + 0.019693277776241302, + 0.020048251375555992, + 0.05715370178222656, + 0.03246723487973213, + 0.051704082638025284, + 0.026352765038609505, + 0.5057951807975769, + 0.15470628440380096 + ], + "credit_map": [ + 0.00927419401705265, + 0.011762107722461224, + 0.02270139381289482, + 0.010265924036502838, + 0.011471383273601532, + 0.015383954159915447, + 0.009618283249437809, + 0.011237199418246746, + 0.005025140475481749, + 0.0029398996848613024, + 0.005764554720371962, + 0.004095706157386303, + 0.0030199545435607433, + 0.0030157817527651787, + 0.0034904987551271915, + 0.0030132902320474386, + 0.0017668387154117227, + 0.0026526530273258686, + 0.0021391138434410095, + 0.0024210086558014154, + 0.005027092061936855, + 0.0012693306198343635, + 0.002613786607980728, + 0.0018034534296020865, + 0.0012582469498738647, + 0.004481108393520117, + 0.0017467144643887877, + 0.0016033785650506616, + 0.005821687169373035, + 0.001397064421325922, + 0.0023049619048833847, + 0.0014350882265716791, + 0.004874343052506447, + 0.008100907318294048, + 0.008494471199810505, + 0.011233041062951088, + 0.005205780733376741, + 0.0037480678874999285, + 0.009166241623461246, + 0.00633084075525403, + 0.002630360424518585, + 0.004401770886033773, + 0.0039948574267327785, + 0.004895725287497044, + 0.0030109796207398176, + 0.002583311405032873, + 0.007737365085631609, + 0.00321285892277956, + 0.0034284901339560747, + 0.0069913859479129314, + 0.005996436811983585, + 0.009167286567389965, + 0.005972991231828928, + 0.003466264344751835, + 0.011877717450261116, + 0.004803502932190895, + 0.0023222635500133038, + 0.005182790104299784, + 0.0023277895525097847, + 0.004225802142173052, + 0.0040001459419727325, + 0.0024954723194241524, + 0.0029273114632815123, + 0.0028711885679513216, + 0.010711906477808952, + 0.06952770799398422, + 0.017190223559737206, + 0.031161848455667496, + 0.19089069962501526, + 0.022087104618549347, + 0.1482129991054535, + 0.01601266674697399, + 0.003482462139800191, + 0.060384541749954224, + 0.0037966871168464422, + 0.009747643023729324, + 0.01588165946304798, + 0.0021423024591058493, + 0.05399369075894356, + 0.005277280695736408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012278917245566845 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0298287533223629 + ], + "opposite_logprob_tokens": [ + -3.5298287868499756 + ], + "logprob_reward": -0.0298287533223629, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298287533223629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2191019207239151, + 0.06825999915599823, + 0.05943695828318596, + 0.044453416019678116, + 0.14016281068325043, + 0.040480151772499084, + 0.02192501910030842, + 0.026319529861211777, + 0.12367662042379379, + 0.25618353486061096 + ], + "credit_map": [ + 0.0227222703397274, + 0.0297174621373415, + 0.031621333211660385, + 0.031486183404922485, + 0.02169094979763031, + 0.028001002967357635, + 0.028361983597278595, + 0.02550072968006134, + 0.008415966294705868, + 0.006809170823544264, + 0.013183358125388622, + 0.008814123459160328, + 0.007985732518136501, + 0.005999163258820772, + 0.0074430652894079685, + 0.009609418921172619, + 0.005932786501944065, + 0.006759242620319128, + 0.006794732064008713, + 0.010480492375791073, + 0.010776020586490631, + 0.005717440042644739, + 0.006807670462876558, + 0.006168571766465902, + 0.003703899448737502, + 0.0073089138604700565, + 0.0057426816783845425, + 0.005697954446077347, + 0.00820798147469759, + 0.003801790764555335, + 0.0053728013299405575, + 0.00461739394813776, + 0.007021937053650618, + 0.014229252934455872, + 0.007468155585229397, + 0.011110830120742321, + 0.04701932147145271, + 0.029555542394518852, + 0.017886077985167503, + 0.005871694069355726, + 0.002590297954156995, + 0.006965918000787497, + 0.004432110581547022, + 0.005519403610378504, + 0.003370806109160185, + 0.0018913382664322853, + 0.01292401272803545, + 0.0027862670831382275, + 0.0014466609572991729, + 0.003541249316185713, + 0.003151470562443137, + 0.0036116503179073334, + 0.003285859478637576, + 0.0023543331772089005, + 0.002642317907884717, + 0.0018914785468950868, + 0.0019631448667496443, + 0.005537146702408791, + 0.0024626762606203556, + 0.003965997602790594, + 0.003678759792819619, + 0.0023947483859956264, + 0.00333992438390851, + 0.0029771325644105673, + 0.005508251953870058, + 0.020196277648210526, + 0.011557470075786114, + 0.012474136427044868, + 0.034927401691675186, + 0.008346443995833397, + 0.020180247724056244, + 0.010486400686204433, + 0.006359391380101442, + 0.09863775223493576, + 0.007288467139005661, + 0.01598804071545601, + 0.026648269966244698, + 0.004671563394367695, + 0.08799908310174942, + 0.008590979501605034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0298287533223629 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597022026777267 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020597022026777267, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597022026777267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07635180652141571, + 0.040822386741638184, + 0.03690458834171295, + 0.038848213851451874, + 0.1273922324180603, + 0.05786507949233055, + 0.08948811143636703, + 0.053937748074531555, + 0.30815935134887695, + 0.1702304631471634 + ], + "credit_map": [ + 0.006067992188036442, + 0.0065601905807852745, + 0.014817038550972939, + 0.011134674772620201, + 0.0092386519536376, + 0.010076923295855522, + 0.01030764915049076, + 0.008148686029016972, + 0.0035944071132689714, + 0.0033888258039951324, + 0.007333292160183191, + 0.005193851888179779, + 0.005637360271066427, + 0.004835152067244053, + 0.006267710123211145, + 0.004571790806949139, + 0.003848925931379199, + 0.004629041999578476, + 0.004066598601639271, + 0.007877164520323277, + 0.005082922987639904, + 0.00270952214486897, + 0.00512783695012331, + 0.0035625763703137636, + 0.001972108380869031, + 0.01229014154523611, + 0.0021008308976888657, + 0.0034301457926630974, + 0.012403788976371288, + 0.0017188630299642682, + 0.0032353666611015797, + 0.0016969661228358746, + 0.008227118290960789, + 0.007556472904980183, + 0.00622281851246953, + 0.011469472199678421, + 0.017211012542247772, + 0.033837758004665375, + 0.03589502349495888, + 0.006972564850002527, + 0.00295056845061481, + 0.016340430825948715, + 0.00578463077545166, + 0.006014661397784948, + 0.005810085218399763, + 0.003422865178436041, + 0.013987569138407707, + 0.003554270137101412, + 0.005429663695394993, + 0.016736257821321487, + 0.007338610477745533, + 0.016116464510560036, + 0.007894395850598812, + 0.004993419162929058, + 0.020220046862959862, + 0.010759251192212105, + 0.002981868339702487, + 0.01199832558631897, + 0.004514401778578758, + 0.008542874827980995, + 0.006515142507851124, + 0.0031388227362185717, + 0.011041958816349506, + 0.005204354878515005, + 0.0072527797892689705, + 0.039107099175453186, + 0.012709831818938255, + 0.023797065019607544, + 0.11165636777877808, + 0.012671430595219135, + 0.09043854475021362, + 0.01052621565759182, + 0.0043600318022072315, + 0.060584813356399536, + 0.005450255703181028, + 0.014291955158114433, + 0.014378855936229229, + 0.0036943729501217604, + 0.059682756662368774, + 0.007787419017404318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.020597022026777267 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978483073413372 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.00978483073413372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978483073413372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10234823822975159, + 0.08161208033561707, + 0.022508515045046806, + 0.013812918215990067, + 0.15989208221435547, + 0.022272661328315735, + 0.03928392380475998, + 0.06722623109817505, + 0.25211772322654724, + 0.23892566561698914 + ], + "credit_map": [ + 0.01386572141200304, + 0.009432081133127213, + 0.015497400425374508, + 0.015573709271848202, + 0.00739976717159152, + 0.012259593233466148, + 0.014417925849556923, + 0.013902035541832447, + 0.010862404480576515, + 0.006516218651086092, + 0.013398165814578533, + 0.010659451596438885, + 0.012069241143763065, + 0.008927485905587673, + 0.010539663024246693, + 0.008639450185000896, + 0.0022179309744387865, + 0.0018776642391458154, + 0.002810919191688299, + 0.0053426106460392475, + 0.0020780342165380716, + 0.0021093501709401608, + 0.0035336920991539955, + 0.0025383152533322573, + 0.0013920495985075831, + 0.002619743812829256, + 0.00143160589504987, + 0.0018803648417815566, + 0.002518339781090617, + 0.0011544154258444905, + 0.0016498015029355884, + 0.0011665961937978864, + 0.008568297140300274, + 0.013905663974583149, + 0.013521864078938961, + 0.018879806622862816, + 0.040558528155088425, + 0.00845717266201973, + 0.046552468091249466, + 0.009448282420635223, + 0.001594775589182973, + 0.0038525424897670746, + 0.0028299621772021055, + 0.002844401868060231, + 0.0019096145406365395, + 0.0021351263858377934, + 0.004497786518186331, + 0.0026084540877491236, + 0.0023691533133387566, + 0.006554103456437588, + 0.004087381064891815, + 0.007600296288728714, + 0.004204747267067432, + 0.0022686184383928776, + 0.007939367555081844, + 0.004260255489498377, + 0.0042946068570017815, + 0.012623812071979046, + 0.006704761181026697, + 0.011795834638178349, + 0.00847899541258812, + 0.005786689929664135, + 0.009041517041623592, + 0.00850001536309719, + 0.0052516888827085495, + 0.03250659629702568, + 0.00802930910140276, + 0.015612456016242504, + 0.09412577748298645, + 0.01124565675854683, + 0.07680068165063858, + 0.008545569144189358, + 0.005850099492818117, + 0.09052559733390808, + 0.006439251825213432, + 0.01649845950305462, + 0.02251179702579975, + 0.0039243148639798164, + 0.08431272208690643, + 0.008863427676260471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.00978483073413372 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343872770667076 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343872770667076, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343872770667076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06060810014605522, + 0.02000977285206318, + 0.006114436779171228, + 0.009863834828138351, + 0.3098312318325043, + 0.007358680013567209, + 0.015908874571323395, + 0.03253249451518059, + 0.4447508156299591, + 0.09302178770303726 + ], + "credit_map": [ + 0.006111296359449625, + 0.006490567233413458, + 0.012626034207642078, + 0.005801315885037184, + 0.005039704032242298, + 0.010063688270747662, + 0.006107347551733255, + 0.008368146605789661, + 0.002515660598874092, + 0.0018701989902183414, + 0.0032769585959613323, + 0.0021575680002570152, + 0.0016478573670610785, + 0.0031925151124596596, + 0.0024729908909648657, + 0.002876023994758725, + 0.0006041954620741308, + 0.0006541524198837578, + 0.0008574455277994275, + 0.0010846513323485851, + 0.0008030635653994977, + 0.00047829924733377993, + 0.0009049244108609855, + 0.0007277050171978772, + 0.0005199503502808511, + 0.0024691580329090357, + 0.0008380849030800164, + 0.0010449732653796673, + 0.0026140885893255472, + 0.0005455928621813655, + 0.0008926622685976326, + 0.0009393246727995574, + 0.016685524955391884, + 0.027985496446490288, + 0.02112431637942791, + 0.03974858298897743, + 0.08691273629665375, + 0.013601182028651237, + 0.09203094244003296, + 0.01174241118133068, + 0.0004627397865988314, + 0.0019462686032056808, + 0.0008219561423175037, + 0.0008257446461357176, + 0.0007862814236432314, + 0.0004090878937859088, + 0.0015628007240593433, + 0.0005438008811324835, + 0.00112216139677912, + 0.0024459981359541416, + 0.0015248063718900084, + 0.003303399309515953, + 0.00240723486058414, + 0.0011333178263157606, + 0.0026101514231413603, + 0.0013618039665743709, + 0.0026904114056378603, + 0.006115133874118328, + 0.0035343742929399014, + 0.0052178543992340565, + 0.004543012008070946, + 0.003390622092410922, + 0.0032635240349918604, + 0.003777563339099288, + 0.009059922769665718, + 0.06063294783234596, + 0.014347528107464314, + 0.025506310164928436, + 0.16927893459796906, + 0.019518814980983734, + 0.13210363686084747, + 0.01430270541459322, + 0.0022716408129781485, + 0.0357685424387455, + 0.0025829339865595102, + 0.005842389538884163, + 0.010254916734993458, + 0.0022179584484547377, + 0.030988803133368492, + 0.0030946016777306795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.026343872770667076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, mesmerizing push-in of a perfectly square block of pale golden butter placed in a dark, seasoned cast iron skillet. Under a warm overhead light, the butter's smooth, waxy surface begins to sweat, its sharp edges softening and collapsing as the heat takes hold. The solid form steadily dissolves from the bottom upward, giving way to an ever-expanding pool of shimmering, translucent liquid gold that catches the light with a brilliant sheen. Tiny, flaky bubbles form where the butter meets the hot pan, gently rising through the melting liquid before gently popping onto the surface. As the camera continues its deliberate forward movement, the last remnants of solid butter boil fully, transforming the scene into a gently bubbling, shimmering pool that ripples hypnotically, filling the frame with its rich, warm colors and creating a deeply satisfying and sensual visual experience.", + "seed": 19790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_559_seed_19790_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": -0.04282849095761776, + "vlm_logprob": -0.04282849095761776, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04282849095761776, + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018850935623049736 + ], + "opposite_logprob_tokens": [ + -4.268850803375244 + ], + "logprob_reward": -0.018850935623049736, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018850935623049736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1475060135126114, + 0.036177411675453186, + 0.05175778642296791, + 0.03770732879638672, + 0.06833860278129578, + 0.1582106649875641, + 0.11446124315261841, + 0.2815335988998413, + 0.025579875335097313, + 0.07872754335403442 + ], + "credit_map": [ + 0.017625683918595314, + 0.00982570368796587, + 0.021161507815122604, + 0.020092172548174858, + 0.014480776153504848, + 0.01770469732582569, + 0.024268480017781258, + 0.0223469827324152, + 0.0047319745644927025, + 0.0028277714736759663, + 0.00449644960463047, + 0.003839281154796481, + 0.003452033968642354, + 0.005142526235431433, + 0.004699416924268007, + 0.006987953092902899, + 0.006619452033191919, + 0.004153007175773382, + 0.008534754626452923, + 0.007580863777548075, + 0.005077772308140993, + 0.004892171826213598, + 0.008050349541008472, + 0.0068494095467031, + 0.004038653336465359, + 0.003355179214850068, + 0.004576214123517275, + 0.007060809060931206, + 0.005289463791996241, + 0.0030074450187385082, + 0.006938169244676828, + 0.0034413933753967285, + 0.00724802864715457, + 0.006466737948358059, + 0.012037323787808418, + 0.006016683764755726, + 0.010959697887301445, + 0.009204389527440071, + 0.007577850949019194, + 0.008827884681522846, + 0.010165523737668991, + 0.01885993592441082, + 0.047835249453783035, + 0.009065789170563221, + 0.020779713988304138, + 0.03465443104505539, + 0.00967335607856512, + 0.007176644634455442, + 0.00726024154573679, + 0.020935239270329475, + 0.029784873127937317, + 0.009145446121692657, + 0.026122888550162315, + 0.007425048854202032, + 0.007253934163600206, + 0.0065335617400705814, + 0.009579657576978207, + 0.1152958869934082, + 0.05460508540272713, + 0.00817704014480114, + 0.06843023002147675, + 0.008501270785927773, + 0.009181159548461437, + 0.007763253524899483, + 0.0015971283428370953, + 0.0037559405900537968, + 0.0038808202371001244, + 0.002903735963627696, + 0.004113806877285242, + 0.0025570273865014315, + 0.004171396140009165, + 0.002600019332021475, + 0.004978790879249573, + 0.007827676832675934, + 0.01951294019818306, + 0.014617894776165485, + 0.008624444715678692, + 0.006207100115716457, + 0.009438191540539265, + 0.00752050057053566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.018850935623049736 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208861619234085 + ], + "opposite_logprob_tokens": [ + -2.81208872795105 + ], + "logprob_reward": -0.06208861619234085, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208861619234085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17012687027454376, + 0.09887560456991196, + 0.0630863606929779, + 0.021175282076001167, + 0.17095331847667694, + 0.11229956895112991, + 0.06119965761899948, + 0.20342586934566498, + 0.04121541231870651, + 0.05764203146100044 + ], + "credit_map": [ + 0.020798970013856888, + 0.01364930346608162, + 0.028024671599268913, + 0.019562389701604843, + 0.01555023156106472, + 0.019519370049238205, + 0.023311728611588478, + 0.02971021458506584, + 0.01529652625322342, + 0.006659925449639559, + 0.016258681192994118, + 0.01131195668131113, + 0.009940728545188904, + 0.012706651352345943, + 0.01023678295314312, + 0.016464347019791603, + 0.006955868098884821, + 0.005587128922343254, + 0.008830287493765354, + 0.010553322732448578, + 0.0066633159294724464, + 0.006495773326605558, + 0.009691464714705944, + 0.008309202268719673, + 0.002965508494526148, + 0.001675268984399736, + 0.0038293623365461826, + 0.0028447380755096674, + 0.00184615736361593, + 0.0019365751650184393, + 0.0034029947128146887, + 0.00267467787489295, + 0.012730701826512814, + 0.020041396841406822, + 0.03438350185751915, + 0.01642187312245369, + 0.040561169385910034, + 0.013409670442342758, + 0.019242744892835617, + 0.014162266626954079, + 0.005087451543658972, + 0.024247100576758385, + 0.012278432957828045, + 0.004098657984286547, + 0.03721368685364723, + 0.02217249944806099, + 0.004333680961281061, + 0.002868066309019923, + 0.004540684632956982, + 0.007972944527864456, + 0.016409052535891533, + 0.005088785197585821, + 0.012791533954441547, + 0.004618768114596605, + 0.005390349309891462, + 0.004387541674077511, + 0.006754525471478701, + 0.08006969839334488, + 0.04395836591720581, + 0.006455004680901766, + 0.04636091738939285, + 0.006046986673027277, + 0.007498530671000481, + 0.0062818522565066814, + 0.003706790041178465, + 0.005383307114243507, + 0.01005961187183857, + 0.00551159493625164, + 0.0038002070505172014, + 0.0030231860000640154, + 0.00393633870407939, + 0.005794378463178873, + 0.0050725494511425495, + 0.006654525175690651, + 0.013368523679673672, + 0.010810350999236107, + 0.005907424259930849, + 0.004533903207629919, + 0.005649311002343893, + 0.0056454455479979515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.06208861619234085 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038139183074235916 + ], + "opposite_logprob_tokens": [ + -3.2881391048431396 + ], + "logprob_reward": -0.038139183074235916, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038139183074235916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05199627950787544, + 0.02217712067067623, + 0.04430536925792694, + 0.0392821803689003, + 0.04267753288149834, + 0.2654970586299896, + 0.01987779140472412, + 0.2968291640281677, + 0.03536655008792877, + 0.18199099600315094 + ], + "credit_map": [ + 0.0060682035982608795, + 0.004572204314172268, + 0.009029722772538662, + 0.006135640665888786, + 0.0046722665429115295, + 0.005355830304324627, + 0.007544775493443012, + 0.008617633022367954, + 0.002179036382585764, + 0.002052412834018469, + 0.0033556020352989435, + 0.003466848749667406, + 0.0022600600495934486, + 0.002618711907416582, + 0.0039363885298371315, + 0.0023080583196133375, + 0.005085494369268417, + 0.0030579024460166693, + 0.0069232177920639515, + 0.0075935604982078075, + 0.00359905487857759, + 0.004138995427638292, + 0.007928849197924137, + 0.005978289991617203, + 0.004855060018599033, + 0.00442121084779501, + 0.005792429205030203, + 0.007763385307043791, + 0.004288517404347658, + 0.0026416077744215727, + 0.0061121778562664986, + 0.003407789161428809, + 0.003418130800127983, + 0.005435074679553509, + 0.006653595250099897, + 0.00313932984136045, + 0.011880072765052319, + 0.004745314363390207, + 0.004265569616109133, + 0.0031404411420226097, + 0.018049011006951332, + 0.029140615835785866, + 0.10840839147567749, + 0.021914424374699593, + 0.015110617503523827, + 0.03093525767326355, + 0.019990066066384315, + 0.02194865047931671, + 0.0017839313950389624, + 0.003917827270925045, + 0.0021682067308574915, + 0.00250009773299098, + 0.0036077930126339197, + 0.0017665235791355371, + 0.0023789934348315, + 0.0017544170841574669, + 0.010089050978422165, + 0.12022491544485092, + 0.06114383414387703, + 0.009274763986468315, + 0.06895329058170319, + 0.009018318727612495, + 0.00964262243360281, + 0.008482350967824459, + 0.002394268289208412, + 0.006458156276494265, + 0.004798846319317818, + 0.003663046518340707, + 0.006083309184759855, + 0.00429091090336442, + 0.004700846038758755, + 0.0029771667905151844, + 0.011935536749660969, + 0.018767615780234337, + 0.04509005323052406, + 0.03902193531394005, + 0.017201857641339302, + 0.013017896562814713, + 0.019631562754511833, + 0.017324522137641907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.038139183074235916 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607857272028923 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.01607857272028923, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607857272028923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05033576115965843, + 0.0188226830214262, + 0.03461281582713127, + 0.0447571724653244, + 0.04272907227277756, + 0.23350508511066437, + 0.044259484857320786, + 0.4413338005542755, + 0.01664491929113865, + 0.07299917936325073 + ], + "credit_map": [ + 0.006051792297512293, + 0.004046602640300989, + 0.009926136583089828, + 0.0068695698864758015, + 0.0036299757193773985, + 0.005842280574142933, + 0.0073836748488247395, + 0.0065857297740876675, + 0.0022398941218852997, + 0.0017210886580869555, + 0.0022940379567444324, + 0.0027119587175548077, + 0.0018158548045903444, + 0.0024915700778365135, + 0.0034044855274260044, + 0.002143793273717165, + 0.0036601817701011896, + 0.002345965476706624, + 0.005517395678907633, + 0.006697372999042273, + 0.0029184981249272823, + 0.0029629168566316366, + 0.0059396130964159966, + 0.004570874851197004, + 0.003952784463763237, + 0.004748454783111811, + 0.006810889113694429, + 0.006570454686880112, + 0.006995508447289467, + 0.0035455473698675632, + 0.007971367798745632, + 0.004162164404988289, + 0.0029512348119169474, + 0.006413064897060394, + 0.006655402481555939, + 0.0023369849659502506, + 0.01378239132463932, + 0.004305935464799404, + 0.0030673188157379627, + 0.0032167399767786264, + 0.013952578417956829, + 0.015047888271510601, + 0.09361809492111206, + 0.016441069543361664, + 0.051825638860464096, + 0.008073525503277779, + 0.01636146940290928, + 0.018184827640652657, + 0.001963701331987977, + 0.01464047934859991, + 0.011349239386618137, + 0.002590393414720893, + 0.007297379896044731, + 0.0016158918151631951, + 0.002711489098146558, + 0.0020909165032207966, + 0.014476645737886429, + 0.17920449376106262, + 0.0901174545288086, + 0.014012574218213558, + 0.10403688997030258, + 0.01290085632354021, + 0.014170929789543152, + 0.012413945980370045, + 0.0012343379203230143, + 0.002114479895681143, + 0.0025947692338377237, + 0.0019906428642570972, + 0.002873519901186228, + 0.0018318075453862548, + 0.0024903437588363886, + 0.0015150184044614434, + 0.004964159801602364, + 0.00791911594569683, + 0.017599331215023994, + 0.014850176870822906, + 0.007944858632981777, + 0.005193864926695824, + 0.007695572450757027, + 0.006832096725702286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.01607857272028923 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898514717817307 + ], + "opposite_logprob_tokens": [ + -2.5789852142333984 + ], + "logprob_reward": -0.07898514717817307, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898514717817307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018334779888391495, + 0.009407514706254005, + 0.02594280242919922, + 0.03241600841283798, + 0.04872601106762886, + 0.23523247241973877, + 0.07419198006391525, + 0.4045121371746063, + 0.027638310566544533, + 0.12359797209501266 + ], + "credit_map": [ + 0.0021594702266156673, + 0.001656153006479144, + 0.0034059504978358746, + 0.0023188902996480465, + 0.0016920024063438177, + 0.002039301209151745, + 0.002406870713457465, + 0.002656139899045229, + 0.0009200789500027895, + 0.0006332411430776119, + 0.0015883703017607331, + 0.001707459450699389, + 0.0007739246357232332, + 0.0010158835211768746, + 0.0015021847793832421, + 0.001266371924430132, + 0.002894242526963353, + 0.001988587900996208, + 0.0039163315668702126, + 0.0029643045272678137, + 0.00178045267239213, + 0.0030557613354176283, + 0.004352907184511423, + 0.0049902149476110935, + 0.0031738821417093277, + 0.0032825206872075796, + 0.004208842758089304, + 0.004780002869665623, + 0.00368491280823946, + 0.003230499802157283, + 0.006351830903440714, + 0.0037035192362964153, + 0.004701628349721432, + 0.0048811789602041245, + 0.011021099053323269, + 0.005973980762064457, + 0.0059495363384485245, + 0.004408331122249365, + 0.005708669777959585, + 0.006081583444029093, + 0.012714197859168053, + 0.03833812102675438, + 0.07596257328987122, + 0.01513356901705265, + 0.0336720310151577, + 0.03095477819442749, + 0.014585715718567371, + 0.01387150026857853, + 0.0030452439095824957, + 0.019241798669099808, + 0.02554967813193798, + 0.0058184522204101086, + 0.0066336458548903465, + 0.0030343015678226948, + 0.0066141062416136265, + 0.004254753235727549, + 0.013682990334928036, + 0.1705160290002823, + 0.069097138941288, + 0.012324552051723003, + 0.10511083900928497, + 0.012281926348805428, + 0.011406674981117249, + 0.010091986507177353, + 0.0019192562904208899, + 0.004574548918753862, + 0.00346878613345325, + 0.0029908171854913235, + 0.004899545572698116, + 0.0036548797506839037, + 0.003796265460550785, + 0.002334211952984333, + 0.007802340667694807, + 0.012392804957926273, + 0.03046264313161373, + 0.02659286931157112, + 0.012585995718836784, + 0.00833649281412363, + 0.013357484713196754, + 0.012067335657775402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.07898514717817307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, mesmerizing push-in of a perfectly square block of pale golden butter placed in a dark, seasoned cast iron skillet. Under a warm overhead light, the butter's smooth, waxy surface begins to sweat, its sharp edges softening and collapsing as the heat takes hold. The solid form steadily dissolves from the bottom upward, giving way to an ever-expanding pool of shimmering, translucent liquid gold that catches the light with a brilliant sheen. Tiny, flaky bubbles form where the butter meets the hot pan, gently rising through the melting liquid before gently popping onto the surface. As the camera continues its deliberate forward movement, the last remnants of solid butter boil fully, transforming the scene into a gently bubbling, shimmering pool that ripples hypnotically, filling the frame with its rich, warm colors and creating a deeply satisfying and sensual visual experience.", + "seed": 19791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_5/index_559_seed_19791_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": -0.026663874043151736, + "vlm_logprob": -0.026663874043151736, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.026663874043151736, + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010074008256196976 + ], + "opposite_logprob_tokens": [ + -5.135074138641357 + ], + "logprob_reward": -0.010074008256196976, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010074008256196976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6191754341125488, + 0.13120798766613007, + 0.06517761945724487, + 0.04404715076088905, + 0.016447653993964195, + 0.012203185819089413, + 0.01685941033065319, + 0.025224845856428146, + 0.042476605623960495, + 0.027180083096027374 + ], + "credit_map": [ + 0.024185750633478165, + 0.12028750777244568, + 0.019068371504545212, + 0.021774617955088615, + 0.24527904391288757, + 0.03152260556817055, + 0.0433565154671669, + 0.11370101571083069, + 0.004533594474196434, + 0.007686642464250326, + 0.004617463331669569, + 0.05159979686141014, + 0.02415604516863823, + 0.004510223399847746, + 0.029111959040164948, + 0.004992249887436628, + 0.0047467490658164024, + 0.005739392712712288, + 0.007206358481198549, + 0.012648310512304306, + 0.005693881306797266, + 0.006137410644441843, + 0.01403252687305212, + 0.008972986601293087, + 0.0037727137096226215, + 0.006019758526235819, + 0.0048915920779109, + 0.004912865348160267, + 0.004975402727723122, + 0.0051907324232161045, + 0.006926387548446655, + 0.007357695139944553, + 0.001787516986951232, + 0.0015749752055853605, + 0.0016539875650778413, + 0.00215597334317863, + 0.00203606765717268, + 0.002251550555229187, + 0.0020808009430766106, + 0.0029067799914628267, + 0.00136936258058995, + 0.0011359181953594089, + 0.0013577592326328158, + 0.0015793723287060857, + 0.001059913425706327, + 0.00173587491735816, + 0.0018616048619151115, + 0.0021033799275755882, + 0.0014950294280424714, + 0.00131532724481076, + 0.002362835919484496, + 0.00408337265253067, + 0.001218798803165555, + 0.0015124535420909524, + 0.0024906687904149294, + 0.0023809242993593216, + 0.0020195322576910257, + 0.0028057817835360765, + 0.004179985728114843, + 0.004889299627393484, + 0.002473135944455862, + 0.002222249982878566, + 0.0031957821920514107, + 0.003439075779169798, + 0.003200279315933585, + 0.004441590514034033, + 0.004821763373911381, + 0.005616429727524519, + 0.004178850911557674, + 0.004127005580812693, + 0.005118999630212784, + 0.01097168494015932, + 0.001934310421347618, + 0.001956476829946041, + 0.005712004378437996, + 0.006944029591977596, + 0.0017617116682231426, + 0.0017107116291299462, + 0.004073282703757286, + 0.003087555058300495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.010074008256196976 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659431710839272 + ], + "opposite_logprob_tokens": [ + -4.758659362792969 + ], + "logprob_reward": -0.008659431710839272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008659431710839272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32803189754486084, + 0.2628539800643921, + 0.10449124127626419, + 0.055528827011585236, + 0.011029082350432873, + 0.04050816595554352, + 0.04748916998505592, + 0.036213140934705734, + 0.06923972815275192, + 0.0446147583425045 + ], + "credit_map": [ + 0.01264125108718872, + 0.06346793472766876, + 0.008985711261630058, + 0.010984307155013084, + 0.12906165421009064, + 0.01719292439520359, + 0.024158315733075142, + 0.061539843678474426, + 0.008968295529484749, + 0.01564350537955761, + 0.008313430473208427, + 0.10378026217222214, + 0.05066046118736267, + 0.00890269223600626, + 0.058746110647916794, + 0.007839231751859188, + 0.008012947626411915, + 0.006398492958396673, + 0.0072833579033613205, + 0.02362777851521969, + 0.005927215330302715, + 0.012664959765970707, + 0.02721570059657097, + 0.01336078904569149, + 0.006875086575746536, + 0.0032377017196267843, + 0.005674807820469141, + 0.011208935640752316, + 0.004072514362633228, + 0.010971778072416782, + 0.007472859229892492, + 0.006015143822878599, + 0.0017816005274653435, + 0.0007884260849095881, + 0.0018306721467524767, + 0.0012182806385681033, + 0.0013959520729258657, + 0.0015613189898431301, + 0.00118715176358819, + 0.0012656793696805835, + 0.004734822548925877, + 0.00264855963177979, + 0.005121714901179075, + 0.008040666580200195, + 0.00429419195279479, + 0.004606423433870077, + 0.006835827603936195, + 0.004225956741720438, + 0.004969357047230005, + 0.0034739405382424593, + 0.007508312352001667, + 0.0075081717222929, + 0.004758776631206274, + 0.007696098648011684, + 0.005981060676276684, + 0.005593449808657169, + 0.003748500719666481, + 0.002739368937909603, + 0.005313720088452101, + 0.006013511214405298, + 0.004040895961225033, + 0.005439247936010361, + 0.004465075675398111, + 0.004452821332961321, + 0.006451742257922888, + 0.006652025040239096, + 0.010624297894537449, + 0.009690179489552975, + 0.008161217905580997, + 0.007229314185678959, + 0.008048475719988346, + 0.012382473796606064, + 0.0038373880088329315, + 0.004018950276076794, + 0.006313796155154705, + 0.009609977714717388, + 0.004487467464059591, + 0.004398070741444826, + 0.006712646223604679, + 0.005236460827291012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008659431710839272 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032031440641731024 + ], + "opposite_logprob_tokens": [ + -5.75320291519165 + ], + "logprob_reward": -0.0032031440641731024, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032031440641731024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4936339557170868, + 0.10249711573123932, + 0.07380838692188263, + 0.04282091557979584, + 0.04259953275322914, + 0.028496475890278816, + 0.02647385373711586, + 0.062558613717556, + 0.07568912953138351, + 0.051421958953142166 + ], + "credit_map": [ + 0.018357569351792336, + 0.09794819355010986, + 0.011633235961198807, + 0.014355722814798355, + 0.1987309455871582, + 0.026856333017349243, + 0.03153403103351593, + 0.094217948615551, + 0.005396614316850901, + 0.006793730892241001, + 0.00660425191745162, + 0.03475228324532509, + 0.01843901351094246, + 0.0048977648839354515, + 0.018174543976783752, + 0.007438917178660631, + 0.006971745286136866, + 0.004562847316265106, + 0.006376024801284075, + 0.012800083495676517, + 0.004784921649843454, + 0.013206272386014462, + 0.01447669230401516, + 0.0106298066675663, + 0.004776338115334511, + 0.004831237252801657, + 0.006044237408787012, + 0.006329790223389864, + 0.003206629538908601, + 0.007388104684650898, + 0.005333983805030584, + 0.004910598509013653, + 0.004629439674317837, + 0.004046070855110884, + 0.00572759285569191, + 0.005114356521517038, + 0.003745101857930422, + 0.0047313435934484005, + 0.006594747770577669, + 0.008010880090296268, + 0.0031951230484992266, + 0.002349280519410968, + 0.003680643392726779, + 0.004135827999562025, + 0.0026900155935436487, + 0.003635626519098878, + 0.0036374712362885475, + 0.005172488745301962, + 0.0032638157717883587, + 0.001925454125739634, + 0.002838742919266224, + 0.004368290305137634, + 0.0024543744511902332, + 0.004182066302746534, + 0.003693685168400407, + 0.0037474273703992367, + 0.0055686854757368565, + 0.005134547594934702, + 0.010302666574716568, + 0.012671934440732002, + 0.005017699673771858, + 0.007798795588314533, + 0.007555412594228983, + 0.008508877828717232, + 0.004201472271233797, + 0.011216954328119755, + 0.006866494193673134, + 0.016728898510336876, + 0.009008004330098629, + 0.005543149542063475, + 0.009900830686092377, + 0.012223332189023495, + 0.003479225793853402, + 0.006113684736192226, + 0.0067399065010249615, + 0.012983418069779873, + 0.004723564255982637, + 0.0037560714408755302, + 0.008123675361275673, + 0.005502417217940092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0032031440641731024 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080872267484665 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080872267484665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080872267484665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33063092827796936, + 0.1781245470046997, + 0.060265157371759415, + 0.029576854780316353, + 0.024155529215931892, + 0.03444276750087738, + 0.04705171659588814, + 0.08244267851114273, + 0.15046361088752747, + 0.06284625083208084 + ], + "credit_map": [ + 0.013776878826320171, + 0.06312181055545807, + 0.013085270300507545, + 0.010833600535988808, + 0.12816327810287476, + 0.018944283947348595, + 0.02136344276368618, + 0.061342328786849976, + 0.0051035149954259396, + 0.008886870928108692, + 0.005642194300889969, + 0.07366818934679031, + 0.029889537021517754, + 0.005492986645549536, + 0.0426323264837265, + 0.006808895152062178, + 0.006032729055732489, + 0.005449640564620495, + 0.006995678413659334, + 0.011880741454660892, + 0.0037185170222073793, + 0.010256209410727024, + 0.0068057565949857235, + 0.009125884622335434, + 0.002153579844161868, + 0.0029997541569173336, + 0.0036956998519599438, + 0.005981292109936476, + 0.0035932203754782677, + 0.0023783044889569283, + 0.00533092999830842, + 0.0034440725576132536, + 0.0025009922683238983, + 0.0023574752267450094, + 0.004024375695735216, + 0.0028491022530943155, + 0.0027810020837932825, + 0.002344285836443305, + 0.003006894141435623, + 0.004291399847716093, + 0.0026569694746285677, + 0.002799658803269267, + 0.005676038097590208, + 0.007548720110207796, + 0.0022119805216789246, + 0.002963847713544965, + 0.005717465188354254, + 0.004868085030466318, + 0.004234859254211187, + 0.002972945338115096, + 0.005468394141644239, + 0.011141788214445114, + 0.0038360573817044497, + 0.005974740255624056, + 0.007984076626598835, + 0.005438850726932287, + 0.007910703308880329, + 0.00787215307354927, + 0.012948598712682724, + 0.015818238258361816, + 0.006844476796686649, + 0.008938404731452465, + 0.009779208339750767, + 0.012330890633165836, + 0.011971138417720795, + 0.014437946490943432, + 0.026008710265159607, + 0.02296796441078186, + 0.016465382650494576, + 0.013890743255615234, + 0.01695983111858368, + 0.027761878445744514, + 0.00581571739166975, + 0.006137595046311617, + 0.011787287890911102, + 0.008762828074395657, + 0.005824231542646885, + 0.008130712434649467, + 0.007259598933160305, + 0.00912828091531992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.011080872267484665 + }, + { + "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 perfectly square block of pale golden butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the butter placed in a dark, seasoned cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the butter melt from the bottom upward, with sharp edges softening and collapsing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny, flaky bubbles form where the butter meets the pan and rise before popping?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, deliberate push-in throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030191391706467 + ], + "opposite_logprob_tokens": [ + -2.35030198097229 + ], + "logprob_reward": -0.10030191391706467, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030191391706467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11445972323417664, + 0.09276306629180908, + 0.13066603243350983, + 0.06223226338624954, + 0.04341214895248413, + 0.072285957634449, + 0.05191440135240555, + 0.045697297900915146, + 0.3576034605503082, + 0.028965596109628677 + ], + "credit_map": [ + 0.008451191708445549, + 0.016935333609580994, + 0.00772902462631464, + 0.008347997441887856, + 0.03290373831987381, + 0.009124167263507843, + 0.011844572611153126, + 0.019123712554574013, + 0.0030474308878183365, + 0.0049763210117816925, + 0.004560627043247223, + 0.036074332892894745, + 0.015471887774765491, + 0.003603593213483691, + 0.020242664963006973, + 0.004786215722560883, + 0.012647602707147598, + 0.010763739235699177, + 0.010419324040412903, + 0.022852323949337006, + 0.008191073313355446, + 0.02258569374680519, + 0.02636059559881687, + 0.01684570126235485, + 0.005619354080408812, + 0.005987186916172504, + 0.0062138792127370834, + 0.008916037157177925, + 0.012828376144170761, + 0.00816029030829668, + 0.007876303978264332, + 0.0066308495588600636, + 0.006476993206888437, + 0.0038361011538654566, + 0.005872612353414297, + 0.005671311169862747, + 0.006600689142942429, + 0.004032359458506107, + 0.005164860747754574, + 0.005757228005677462, + 0.007082852069288492, + 0.005432574078440666, + 0.009695881977677345, + 0.011876326985657215, + 0.008413969539105892, + 0.010873300023376942, + 0.009042974561452866, + 0.009868089109659195, + 0.004686528351157904, + 0.004777171649038792, + 0.007478593848645687, + 0.009399143047630787, + 0.003690373385325074, + 0.004641510546207428, + 0.007499403785914183, + 0.009741682559251785, + 0.00450023403391242, + 0.005468530114740133, + 0.007650434970855713, + 0.009334051050245762, + 0.003894823370501399, + 0.003218594240024686, + 0.005359522067010403, + 0.006271109450608492, + 0.023192856460809708, + 0.03160649165511131, + 0.052719417959451675, + 0.10541408509016037, + 0.02744402177631855, + 0.020151833072304726, + 0.057036127895116806, + 0.040038660168647766, + 0.002314841141924262, + 0.0028479706961661577, + 0.004796613473445177, + 0.005771972704678774, + 0.002844511764124036, + 0.0027207706589251757, + 0.004109117202460766, + 0.003559800097718835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.10030191391706467 + } + ], + "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/0000017/group_6/index_tensor([1286, 1286])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_tensor([1286, 1286])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..500d2a762712a3aed4ae7148db4119768cb1616b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_tensor([1286, 1286])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in dramatic slow motion with a static camera, the chain reaction is meticulously documented in a surreal, extreme close-up. A pristine, unmarked white aluminum aerosol can with a matte finish stands atop a dark, polished wood surface, its shape defined by intense, cool sidelighting, casting long shadows. The scene begins with low-frequency vibrations subtly vibrating the surface, causing the can to tremble almost imperceptibly. As the vibrations intensify, the can begins to tilt inexorably, its base slowly losing its grip on the smooth wood. It topples in a graceful arc, its white plastic actuator cap striking the hard edge of the wooden surface with a sharp, silent impact. The force of the collision instantly compresses the actuator, forcing a powerful, cone-shaped plume of fine mist from the nozzle. The mist billows outward, its individual particles catching the dramatic sidelight, forming a shimmering, ethereal cloud against the dark, out-of-focus background, showcasing the mechanics with clinical and stunning visual clarity.", + "seed": 7807, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7807_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": -0.33699601646512745, + "vlm_logprob": -0.33699601646512745, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.33699601646512745, + "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 video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18812797963619232 + ], + "opposite_logprob_tokens": [ + -2.0631279945373535 + ], + "logprob_reward": -0.18812797963619232, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.18812797963619232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06452473253011703, + 0.06018417328596115, + 0.06836766749620438, + 0.08133444935083389, + 0.13416068255901337, + 0.12261239439249039, + 0.0671088770031929, + 0.11537499725818634, + 0.20339274406433105, + 0.0829392746090889 + ], + "credit_map": [ + 0.009911536239087582, + 0.006159938406199217, + 0.009437876753509045, + 0.00912562757730484, + 0.0077206408604979515, + 0.00914083980023861, + 0.006781192496418953, + 0.006247082259505987, + 0.008048790507018566, + 0.008927149698138237, + 0.0070083774626255035, + 0.007912878878414631, + 0.007086127996444702, + 0.008645611815154552, + 0.006536000873893499, + 0.006019240245223045, + 0.011960415169596672, + 0.005501470062881708, + 0.016819950193166733, + 0.014645940624177456, + 0.003933956380933523, + 0.004270550329238176, + 0.00568270031362772, + 0.005552692338824272, + 0.006501535419374704, + 0.008019626140594482, + 0.011705721728503704, + 0.02667197212576866, + 0.005373471416532993, + 0.004101011902093887, + 0.013446460478007793, + 0.005514655262231827, + 0.017139017581939697, + 0.019196338951587677, + 0.024583850055933, + 0.0182830598205328, + 0.010986832901835442, + 0.017017725855112076, + 0.01896708831191063, + 0.007986776530742645, + 0.011932813562452793, + 0.014944279566407204, + 0.02002231776714325, + 0.028388459235429764, + 0.009831344708800316, + 0.012272689491510391, + 0.017353639006614685, + 0.00786686036735773, + 0.005263404920697212, + 0.01212692353874445, + 0.009478497318923473, + 0.00959084928035736, + 0.010336478240787983, + 0.00636718887835741, + 0.009137879125773907, + 0.004807662684470415, + 0.014218715019524097, + 0.01768440194427967, + 0.016221623867750168, + 0.014796888455748558, + 0.022968515753746033, + 0.008325958624482155, + 0.015048490837216377, + 0.006110411137342453, + 0.008416465483605862, + 0.06899992376565933, + 0.05277961120009422, + 0.017272857949137688, + 0.03350493684411049, + 0.006645714398473501, + 0.009324262849986553, + 0.006448986008763313, + 0.0040270560421049595, + 0.02849232777953148, + 0.022191394120454788, + 0.005989945959299803, + 0.012457948178052902, + 0.0025454082060605288, + 0.004079860635101795, + 0.0031553409062325954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.18812797963619232 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.4063507318496704 + ], + "opposite_logprob_tokens": [ + -0.2813507616519928 + ], + "logprob_reward": -1.4063507318496704, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063507318496704, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06071373075246811, + 0.06627465039491653, + 0.05999591201543808, + 0.059964895248413086, + 0.060144826769828796, + 0.07807597517967224, + 0.09758099168539047, + 0.02540559694170952, + 0.04547485336661339, + 0.4463685154914856 + ], + "credit_map": [ + 0.0075452993623912334, + 0.006353465840220451, + 0.006600765977054834, + 0.009795946069061756, + 0.006548580713570118, + 0.007083415985107422, + 0.01031105313450098, + 0.006475207395851612, + 0.009233425371348858, + 0.006296439096331596, + 0.007782553788274527, + 0.007538148667663336, + 0.008841601200401783, + 0.008236035704612732, + 0.010796168819069862, + 0.007550284266471863, + 0.007565758191049099, + 0.007050350774079561, + 0.008182861842215061, + 0.008779993280768394, + 0.006302343215793371, + 0.00588252255693078, + 0.009625573642551899, + 0.006606516893953085, + 0.004989979322999716, + 0.006431959103792906, + 0.011650990694761276, + 0.013946639373898506, + 0.0041667139157652855, + 0.005877090618014336, + 0.007261727936565876, + 0.0056397998705506325, + 0.008692448027431965, + 0.006392177194356918, + 0.016737764701247215, + 0.00905524380505085, + 0.0032865917310118675, + 0.005213180091232061, + 0.00794999860227108, + 0.0028174258768558502, + 0.006661910098046064, + 0.007451200857758522, + 0.015616691671311855, + 0.022024329751729965, + 0.005163456313312054, + 0.005038934294134378, + 0.01131480187177658, + 0.004804653115570545, + 0.010754681192338467, + 0.018037546426057816, + 0.012658982537686825, + 0.014322665520012379, + 0.021931404247879982, + 0.0053695198148489, + 0.009890930727124214, + 0.004615260753780603, + 0.002051933668553829, + 0.0026652077212929726, + 0.003278193064033985, + 0.004955282434821129, + 0.003576930146664381, + 0.002105112886056304, + 0.004320627544075251, + 0.002452310873195529, + 0.0025047515518963337, + 0.013666311278939247, + 0.006770633161067963, + 0.005211018957197666, + 0.0083299670368433, + 0.0024721352383494377, + 0.003918883856385946, + 0.002601157408207655, + 0.019259236752986908, + 0.15429924428462982, + 0.13163834810256958, + 0.03774993121623993, + 0.06175854802131653, + 0.011723385192453861, + 0.019967561587691307, + 0.009972273372113705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -1.4063507318496704 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014232791028916836 + ], + "opposite_logprob_tokens": [ + -4.264232635498047 + ], + "logprob_reward": -0.014232791028916836, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014232791028916836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045863062143325806, + 0.035897769033908844, + 0.04880331829190254, + 0.08965420722961426, + 0.05689981207251549, + 0.12310284376144409, + 0.10655060410499573, + 0.06120571866631508, + 0.07339742034673691, + 0.3586251735687256 + ], + "credit_map": [ + 0.005692238919436932, + 0.004986216314136982, + 0.004745757207274437, + 0.008335568010807037, + 0.005140757653862238, + 0.005338561255484819, + 0.006238337140530348, + 0.005385626573115587, + 0.00495834369212389, + 0.005110808182507753, + 0.004448262043297291, + 0.004944604355841875, + 0.003590268548578024, + 0.0045776404440402985, + 0.004687841050326824, + 0.003580001648515463, + 0.006446266081184149, + 0.005699499510228634, + 0.006509384606033564, + 0.006094662938266993, + 0.0049729556776583195, + 0.00552564300596714, + 0.007970568723976612, + 0.005584335420280695, + 0.00889455433934927, + 0.007741210050880909, + 0.016849424690008163, + 0.016402216628193855, + 0.00900935661047697, + 0.007965391501784325, + 0.013169802725315094, + 0.009622249752283096, + 0.004019182175397873, + 0.007179517764598131, + 0.00892559438943863, + 0.007310221903026104, + 0.006341692991554737, + 0.008366919122636318, + 0.010331192053854465, + 0.004425496328622103, + 0.010733231902122498, + 0.02313639037311077, + 0.019974762573838234, + 0.018127698451280594, + 0.012041952461004257, + 0.016226710751652718, + 0.012932143174111843, + 0.009929957799613476, + 0.008455406874418259, + 0.01707802712917328, + 0.016212182119488716, + 0.01920768991112709, + 0.011278411373496056, + 0.00808247085660696, + 0.017298832535743713, + 0.008937578648328781, + 0.006104845553636551, + 0.010376852005720139, + 0.008355866186320782, + 0.009408663958311081, + 0.009878958575427532, + 0.004057334735989571, + 0.008427857421338558, + 0.004595342557877302, + 0.006996380165219307, + 0.010520324110984802, + 0.016919514164328575, + 0.010073608718812466, + 0.009506545029580593, + 0.00582525972276926, + 0.009144095703959465, + 0.004411695059388876, + 0.015818800777196884, + 0.12853264808654785, + 0.0948077142238617, + 0.029696494340896606, + 0.05879078805446625, + 0.007020397577434778, + 0.015211272984743118, + 0.008747071959078312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.014232791028916836 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207069382071495 + ], + "opposite_logprob_tokens": [ + -2.812070608139038 + ], + "logprob_reward": -0.06207069382071495, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207069382071495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06361265480518341, + 0.037802740931510925, + 0.055858444422483444, + 0.3422108292579651, + 0.07914929836988449, + 0.1325768679380417, + 0.17682243883609772, + 0.03625308349728584, + 0.054224658757448196, + 0.021489035338163376 + ], + "credit_map": [ + 0.008804742246866226, + 0.005273879040032625, + 0.007545263506472111, + 0.008872322738170624, + 0.007096332963556051, + 0.007546191569417715, + 0.006522173993289471, + 0.011951735243201256, + 0.006179672200232744, + 0.005064958240836859, + 0.005479492712765932, + 0.003997310064733028, + 0.0036963748279958963, + 0.00464547798037529, + 0.005038714036345482, + 0.003700736677274108, + 0.008262070827186108, + 0.008314043283462524, + 0.008926472626626492, + 0.008408024907112122, + 0.004602140747010708, + 0.005414763931185007, + 0.006255378481000662, + 0.005675541236996651, + 0.024844884872436523, + 0.026610491797327995, + 0.06993313133716583, + 0.1186489686369896, + 0.01492904219776392, + 0.015674997121095657, + 0.05658554658293724, + 0.014983734115958214, + 0.006887036841362715, + 0.016121290624141693, + 0.010259256698191166, + 0.009507660754024982, + 0.010050324723124504, + 0.010502866469323635, + 0.00970526784658432, + 0.006115589290857315, + 0.011816533282399178, + 0.021844154223799706, + 0.022084327414631844, + 0.021136058494448662, + 0.01276127528399229, + 0.013893861323595047, + 0.01980954222381115, + 0.009231089614331722, + 0.02050795964896679, + 0.03440891206264496, + 0.022926636040210724, + 0.02606147527694702, + 0.039179280400276184, + 0.008169532753527164, + 0.01837974041700363, + 0.0071888938546180725, + 0.003406234085559845, + 0.0032280192244797945, + 0.004664137493818998, + 0.005227668676525354, + 0.005830016452819109, + 0.0030725563410669565, + 0.006567099597305059, + 0.004257350228726864, + 0.004132479894906282, + 0.01206320058554411, + 0.013280848041176796, + 0.0076381550170481205, + 0.006025846116244793, + 0.0023093558847904205, + 0.006181662902235985, + 0.0025931047275662422, + 0.0014623928582295775, + 0.0029431248549371958, + 0.0025738475378602743, + 0.0033181768376380205, + 0.003164128866046667, + 0.002509982557967305, + 0.0026226327754557133, + 0.0028947461396455765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.06207069382071495 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08497273921966553, + 0.07895217835903168, + 0.07465209811925888, + 0.09786243736743927, + 0.12002377957105637, + 0.17149119079113007, + 0.1048893854022026, + 0.10207235813140869, + 0.04730387404561043, + 0.1177799254655838 + ], + "credit_map": [ + 0.013365319930016994, + 0.0068771010264754295, + 0.012611690908670425, + 0.011878296732902527, + 0.010879425331950188, + 0.008948276750743389, + 0.010286172851920128, + 0.010126453824341297, + 0.010256977751851082, + 0.010679632425308228, + 0.008392104879021645, + 0.012879687361419201, + 0.007027621380984783, + 0.008119063451886177, + 0.012251581065356731, + 0.009345509111881256, + 0.016085779294371605, + 0.007078769151121378, + 0.01641441322863102, + 0.01300683431327343, + 0.004072944168001413, + 0.005791539326310158, + 0.005347422789782286, + 0.006854394916445017, + 0.00838837306946516, + 0.007589282467961311, + 0.021726038306951523, + 0.023541253060102463, + 0.006597285158932209, + 0.00845765694975853, + 0.013375597074627876, + 0.008186952210962772, + 0.011800271458923817, + 0.0143709322437644, + 0.016470788046717644, + 0.01480268593877554, + 0.0121232308447361, + 0.024974627420306206, + 0.01516556367278099, + 0.010315683670341969, + 0.014862651005387306, + 0.02300656959414482, + 0.02934388257563114, + 0.0367349348962307, + 0.014269856736063957, + 0.01765718124806881, + 0.024129826575517654, + 0.01148628443479538, + 0.009777596220374107, + 0.022223759442567825, + 0.014095496386289597, + 0.015220069326460361, + 0.01831946149468422, + 0.006877079140394926, + 0.013747829012572765, + 0.004628089722245932, + 0.013098296709358692, + 0.01485690288245678, + 0.012044081464409828, + 0.014295968227088451, + 0.01858464628458023, + 0.00803763885051012, + 0.014213429763913155, + 0.006941392086446285, + 0.004185041878372431, + 0.012557772919535637, + 0.005028391256928444, + 0.005161952693015337, + 0.007625142578035593, + 0.004012802615761757, + 0.005106633063405752, + 0.0036261354107409716, + 0.006177116651087999, + 0.03279080241918564, + 0.037074778228998184, + 0.013286505825817585, + 0.010296745225787163, + 0.006760972086340189, + 0.006987856235355139, + 0.004405156243592501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.014197885990142822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion with a static camera, the chain reaction is meticulously documented in a surreal, extreme close-up. A pristine, unmarked white aluminum aerosol can with a matte finish stands atop a dark, polished wood surface, its shape defined by intense, cool sidelighting, casting long shadows. The scene begins with low-frequency vibrations subtly vibrating the surface, causing the can to tremble almost imperceptibly. As the vibrations intensify, the can begins to tilt inexorably, its base slowly losing its grip on the smooth wood. It topples in a graceful arc, its white plastic actuator cap striking the hard edge of the wooden surface with a sharp, silent impact. The force of the collision instantly compresses the actuator, forcing a powerful, cone-shaped plume of fine mist from the nozzle. The mist billows outward, its individual particles catching the dramatic sidelight, forming a shimmering, ethereal cloud against the dark, out-of-focus background, showcasing the mechanics with clinical and stunning visual clarity.", + "seed": 7808, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7808_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": -0.3133702932856977, + "vlm_logprob": -0.3133702932856977, + "vlm_yes_no_margin": 2.724999988079071, + "vlm_token_credit": -0.3133702932856977, + "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 video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.192844957113266 + ], + "opposite_logprob_tokens": [ + -2.192844867706299 + ], + "logprob_reward": -0.192844957113266, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.192844957113266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5847051739692688, + 0.017774775624275208, + 0.03765278682112694, + 0.038835376501083374, + 0.04157128185033798, + 0.061351239681243896, + 0.07557305693626404, + 0.04692573472857475, + 0.029730642214417458, + 0.06587991118431091 + ], + "credit_map": [ + 0.03288496285676956, + 0.14753392338752747, + 0.05721132829785347, + 0.23763540387153625, + 0.018261993303894997, + 0.018425356596708298, + 0.06304610520601273, + 0.009706120006740093, + 0.0027143466286361217, + 0.002230563201010227, + 0.0023256021086126566, + 0.0038375696167349815, + 0.0016881234478205442, + 0.0014236369170248508, + 0.0019100855570286512, + 0.0016448491951450706, + 0.004932200536131859, + 0.006574887782335281, + 0.0033518881537020206, + 0.0065293037332594395, + 0.005048907361924648, + 0.0034862631000578403, + 0.003678383305668831, + 0.004050948657095432, + 0.0035570983309298754, + 0.0034603888634592295, + 0.003374019404873252, + 0.004703477490693331, + 0.00821805652230978, + 0.0054979571141302586, + 0.005837725009769201, + 0.004186656326055527, + 0.003968085162341595, + 0.004683909937739372, + 0.0037572321016341448, + 0.0045054592192173, + 0.007428291719406843, + 0.005938020069152117, + 0.006461085751652718, + 0.004829195328056812, + 0.004405993036925793, + 0.0075092320330441, + 0.00543120177462697, + 0.008598131127655506, + 0.006343914661556482, + 0.009015245363116264, + 0.014327313750982285, + 0.005720208398997784, + 0.006952097173780203, + 0.008490199223160744, + 0.008251155726611614, + 0.010951962321996689, + 0.009638005867600441, + 0.008374934084713459, + 0.013905891217291355, + 0.009008805267512798, + 0.003270039102062583, + 0.003505474189296365, + 0.004136307165026665, + 0.005220642313361168, + 0.0051389215514063835, + 0.0053903087973594666, + 0.013061877340078354, + 0.007202164735645056, + 0.0020303211640566587, + 0.003412003396078944, + 0.00254750344902277, + 0.004772531799972057, + 0.0033050812780857086, + 0.005475425627082586, + 0.003147150855511427, + 0.005040626507252455, + 0.004999537952244282, + 0.009628132916986942, + 0.007894739508628845, + 0.009408731944859028, + 0.01048233825713396, + 0.005166183691471815, + 0.010574058629572392, + 0.007726188283413649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.192844957113266 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.3134737014770508 + ], + "opposite_logprob_tokens": [ + -0.31347376108169556 + ], + "logprob_reward": -1.3134737014770508, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3134737014770508, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.36608847975730896, + 0.02992974780499935, + 0.0678492859005928, + 0.05905773118138313, + 0.05099481716752052, + 0.032699789851903915, + 0.17085298895835876, + 0.07459987699985504, + 0.029755033552646637, + 0.11817222088575363 + ], + "credit_map": [ + 0.02017112821340561, + 0.08953920006752014, + 0.03843173757195473, + 0.1495966911315918, + 0.01058442983776331, + 0.011785119771957397, + 0.03971121460199356, + 0.006268935743719339, + 0.004117522854357958, + 0.0038146628066897392, + 0.0037099483888596296, + 0.00534904096275568, + 0.0024962332099676132, + 0.003215608885511756, + 0.0035547048319131136, + 0.003672028658911586, + 0.007329646032303572, + 0.013577619567513466, + 0.005497086327522993, + 0.007884165272116661, + 0.00986824557185173, + 0.005627776030451059, + 0.0068006799556314945, + 0.011264066211879253, + 0.0073367212899029255, + 0.008211001753807068, + 0.0060445466078817844, + 0.006679540034383535, + 0.012342246249318123, + 0.0075674536637961864, + 0.006858257111161947, + 0.004017963539808989, + 0.004600584041327238, + 0.00497237965464592, + 0.006087056361138821, + 0.007175028324127197, + 0.007037832867354155, + 0.007020128890872002, + 0.009225516580045223, + 0.00487629184499383, + 0.0022373953834176064, + 0.003919118549674749, + 0.0029683257453143597, + 0.0037591124419122934, + 0.0038182693533599377, + 0.005135819781571627, + 0.00781959854066372, + 0.0030421526171267033, + 0.018679695203900337, + 0.01728127710521221, + 0.029023272916674614, + 0.03047465905547142, + 0.014472694136202335, + 0.01682399958372116, + 0.026725808158516884, + 0.01737157627940178, + 0.008758332580327988, + 0.005739394575357437, + 0.014945815317332745, + 0.010459067299962044, + 0.006675959564745426, + 0.0069913966581225395, + 0.011147163808345795, + 0.009882749058306217, + 0.002480393508449197, + 0.0028229013551026583, + 0.0027040671557188034, + 0.004067155998200178, + 0.0038459300994873047, + 0.0047097825445234776, + 0.0041559236124157906, + 0.004968880210071802, + 0.009214640595018864, + 0.014214521273970604, + 0.01523491833359003, + 0.026173090562224388, + 0.014297151938080788, + 0.007728333584964275, + 0.017281612381339073, + 0.014027959667146206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -1.3134737014770508 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076600625179708 + ], + "opposite_logprob_tokens": [ + -4.882659912109375 + ], + "logprob_reward": -0.0076600625179708, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076600625179708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6587072014808655, + 0.037572480738162994, + 0.030424704775214195, + 0.035050779581069946, + 0.0484217070043087, + 0.03218858316540718, + 0.08704549819231033, + 0.015685822814702988, + 0.028087638318538666, + 0.026815632358193398 + ], + "credit_map": [ + 0.03623463958501816, + 0.16300112009048462, + 0.06923063099384308, + 0.2675500512123108, + 0.019706593826413155, + 0.020728055387735367, + 0.07122498005628586, + 0.011031142435967922, + 0.005164331290870905, + 0.006616060622036457, + 0.005125378258526325, + 0.006060423795133829, + 0.0035516656935214996, + 0.0032996805384755135, + 0.00449565751478076, + 0.0032592813950031996, + 0.003511978778988123, + 0.00510833878070116, + 0.0031488067470490932, + 0.0033142331521958113, + 0.00385281746275723, + 0.0025430582463741302, + 0.003675210988149047, + 0.005270261783152819, + 0.003585945814847946, + 0.004776193294674158, + 0.003023394150659442, + 0.004159830044955015, + 0.0060478379018604755, + 0.004344049375504255, + 0.0051920171827077866, + 0.003921511583030224, + 0.004318479914218187, + 0.0050543914549052715, + 0.004497361369431019, + 0.00571769243106246, + 0.007061666809022427, + 0.007189252879470587, + 0.010466914623975754, + 0.004115948919206858, + 0.0021768996957689524, + 0.0029674640391021967, + 0.0027566305361688137, + 0.004464413970708847, + 0.0031054129358381033, + 0.005277826450765133, + 0.008179055526852608, + 0.0032608809415251017, + 0.008817470632493496, + 0.008927922695875168, + 0.01393663976341486, + 0.015400141477584839, + 0.007333476096391678, + 0.009118400514125824, + 0.015127773396670818, + 0.008383670821785927, + 0.001175682875327766, + 0.0011551865609362721, + 0.0013541735243052244, + 0.0021701729856431484, + 0.0015646216925233603, + 0.0021077096462249756, + 0.0032338539604097605, + 0.0029244208708405495, + 0.0024726095143705606, + 0.002934536198154092, + 0.0025408451911062002, + 0.004137828014791012, + 0.0036393594928085804, + 0.0030586090870201588, + 0.005270506255328655, + 0.004033347591757774, + 0.0019016234437003732, + 0.0035889737773686647, + 0.00325788464397192, + 0.0066228099167346954, + 0.003091755323112011, + 0.0018582912161946297, + 0.003058583475649357, + 0.003435710212215781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0076600625179708 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308897629380226 + ], + "opposite_logprob_tokens": [ + -3.168088912963867 + ], + "logprob_reward": -0.04308897629380226, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308897629380226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8316366076469421, + 0.0082204919308424, + 0.024861503392457962, + 0.0371309332549572, + 0.018107200041413307, + 0.014286024495959282, + 0.01530272513628006, + 0.01977984420955181, + 0.016898097470402718, + 0.013776577077805996 + ], + "credit_map": [ + 0.04671524837613106, + 0.20785249769687653, + 0.08528002351522446, + 0.3379828929901123, + 0.025602899491786957, + 0.025998638942837715, + 0.08873952180147171, + 0.013464927673339844, + 0.0009725947747938335, + 0.0011038911761716008, + 0.0010454196017235518, + 0.0018165969522669911, + 0.0007751886150799692, + 0.0005693595740012825, + 0.0011948549654334784, + 0.0007425862131640315, + 0.002583715133368969, + 0.0030763125978410244, + 0.002177224727347493, + 0.0032803919166326523, + 0.003963186405599117, + 0.0022848064545542, + 0.003037888091057539, + 0.0044579776003956795, + 0.006627910770475864, + 0.0031838323920965195, + 0.0074205221608281136, + 0.005003643222153187, + 0.003409548895433545, + 0.003496776334941387, + 0.004009628668427467, + 0.003979068715125322, + 0.0014341367641463876, + 0.0020051023457199335, + 0.002007989911362529, + 0.00234748562797904, + 0.00259762117639184, + 0.00236352882348001, + 0.00310500874184072, + 0.002246326766908169, + 0.001443760935217142, + 0.002245106268674135, + 0.0013533168239519, + 0.0023201305884867907, + 0.0018314971821382642, + 0.001857694354839623, + 0.0017758053727447987, + 0.0014587127370759845, + 0.0016146621201187372, + 0.0016075129387900233, + 0.002407281193882227, + 0.0022750645875930786, + 0.0014801020734012127, + 0.0017030377639457583, + 0.0027455277740955353, + 0.0014695359859615564, + 0.0011338546173647046, + 0.0013151232851669192, + 0.0016670856857672334, + 0.0030770297162234783, + 0.0020897635258734226, + 0.002088953973725438, + 0.006094401702284813, + 0.002313632518053055, + 0.0013433676213026047, + 0.0013473002472892404, + 0.0012191243004053831, + 0.003649352816864848, + 0.0018528122454881668, + 0.001571323606185615, + 0.0024363985285162926, + 0.003478419268503785, + 0.0010584243573248386, + 0.0015014507807791233, + 0.0018088433425873518, + 0.0033793000038713217, + 0.0014326601522043347, + 0.0009113198611885309, + 0.0021199288312345743, + 0.001564650796353817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.04308897629380226 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8136632442474365, + 0.007734979037195444, + 0.015005617402493954, + 0.021618694067001343, + 0.019335446879267693, + 0.015862220898270607, + 0.03232581913471222, + 0.03464355319738388, + 0.017728861421346664, + 0.022081520408391953 + ], + "credit_map": [ + 0.04594790190458298, + 0.2041492760181427, + 0.08299131691455841, + 0.33029451966285706, + 0.02514507994055748, + 0.02520781010389328, + 0.08676820248365402, + 0.013159236870706081, + 0.0010351785458624363, + 0.0009809542680159211, + 0.0011067542945966125, + 0.0014025097480043769, + 0.0009243927197530866, + 0.0006111820111982524, + 0.0009908390929922462, + 0.0006831695209257305, + 0.001755937933921814, + 0.002348311711102724, + 0.0012610011035576463, + 0.0017127771861851215, + 0.0018925723852589726, + 0.0011371792061254382, + 0.0015781818656250834, + 0.0033196581061929464, + 0.001952574704773724, + 0.002595660975202918, + 0.0019186697900295258, + 0.002982013626024127, + 0.003938478883355856, + 0.002842271700501442, + 0.002723998622968793, + 0.00266502914018929, + 0.0015474811661988497, + 0.0019612538162618876, + 0.0019397442229092121, + 0.002266358584165573, + 0.0032170622143894434, + 0.0028655482456088066, + 0.004222117830067873, + 0.0013158852234482765, + 0.001333659398369491, + 0.002569043543189764, + 0.0016095663886517286, + 0.0024228731635957956, + 0.0017763470532372594, + 0.0021311999298632145, + 0.002296918071806431, + 0.0017226157942786813, + 0.00375028676353395, + 0.0033729004207998514, + 0.005795084405690432, + 0.005807222332805395, + 0.0030796783976256847, + 0.003105663228780031, + 0.004393195267766714, + 0.003021793905645609, + 0.0018654989544302225, + 0.002315133111551404, + 0.0023816353641450405, + 0.0051198601722717285, + 0.004380233120173216, + 0.003652451327070594, + 0.011100788600742817, + 0.0038279560394585133, + 0.0013618370285257697, + 0.0015849468763917685, + 0.0012174618896096945, + 0.003778912825509906, + 0.0020793697331100702, + 0.00174233119469136, + 0.0025340558495372534, + 0.003429946955293417, + 0.0016831791726872325, + 0.0028817455749958754, + 0.0027285623364150524, + 0.005049817729741335, + 0.002701552351936698, + 0.0015583762433379889, + 0.0029659734573215246, + 0.0025123159866780043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.009783769026398659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion with a static camera, the chain reaction is meticulously documented in a surreal, extreme close-up. A pristine, unmarked white aluminum aerosol can with a matte finish stands atop a dark, polished wood surface, its shape defined by intense, cool sidelighting, casting long shadows. The scene begins with low-frequency vibrations subtly vibrating the surface, causing the can to tremble almost imperceptibly. As the vibrations intensify, the can begins to tilt inexorably, its base slowly losing its grip on the smooth wood. It topples in a graceful arc, its white plastic actuator cap striking the hard edge of the wooden surface with a sharp, silent impact. The force of the collision instantly compresses the actuator, forcing a powerful, cone-shaped plume of fine mist from the nozzle. The mist billows outward, its individual particles catching the dramatic sidelight, forming a shimmering, ethereal cloud against the dark, out-of-focus background, showcasing the mechanics with clinical and stunning visual clarity.", + "seed": 7809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7809_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": -0.4403538558632135, + "vlm_logprob": -0.4403538558632135, + "vlm_yes_no_margin": 1.750000011920929, + "vlm_token_credit": -0.4403538558632135, + "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 video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.50264972448349 + ], + "opposite_logprob_tokens": [ + -1.2526497840881348 + ], + "logprob_reward": -0.50264972448349, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.50264972448349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07179225981235504, + 0.1372392773628235, + 0.14594919979572296, + 0.11355697363615036, + 0.18020310997962952, + 0.07353723049163818, + 0.14751923084259033, + 0.03466811403632164, + 0.045655880123376846, + 0.04987872391939163 + ], + "credit_map": [ + 0.011062311939895153, + 0.007757191080600023, + 0.008420911617577076, + 0.009357680566608906, + 0.006349456496536732, + 0.00844046100974083, + 0.008717440068721771, + 0.011686807498335838, + 0.016199972480535507, + 0.01945103518664837, + 0.019286517053842545, + 0.027896815910935402, + 0.013185814023017883, + 0.013560066930949688, + 0.014138179831206799, + 0.013520874083042145, + 0.014496193267405033, + 0.021265221759676933, + 0.020461848005652428, + 0.050491590052843094, + 0.009095770306885242, + 0.007213430479168892, + 0.015317119657993317, + 0.0076080248691141605, + 0.009332282468676567, + 0.010095199570059776, + 0.013007710687816143, + 0.027907315641641617, + 0.015719925984740257, + 0.013569550588726997, + 0.012157215736806393, + 0.011767777614295483, + 0.02339625358581543, + 0.03533979505300522, + 0.018088506534695625, + 0.021505430340766907, + 0.01964089833199978, + 0.02256663702428341, + 0.01705886982381344, + 0.022606704384088516, + 0.008947103284299374, + 0.007268321234732866, + 0.010287659242749214, + 0.01218722015619278, + 0.008270896971225739, + 0.00648431945592165, + 0.009390067309141159, + 0.01070164330303669, + 0.009706808254122734, + 0.0385315828025341, + 0.010584822855889797, + 0.014462621882557869, + 0.04663192853331566, + 0.011293097399175167, + 0.009363465011119843, + 0.006944903172552586, + 0.0035097587388008833, + 0.00458162184804678, + 0.0029413343872874975, + 0.0055222888477146626, + 0.005056142341345549, + 0.0034712336491793394, + 0.004090489819645882, + 0.005495244637131691, + 0.004719363059848547, + 0.007975488901138306, + 0.011174988932907581, + 0.005414116661995649, + 0.005572741385549307, + 0.004492185078561306, + 0.0030170727986842394, + 0.003289925865828991, + 0.0071548172272741795, + 0.006966863293200731, + 0.011626512743532658, + 0.007019209675490856, + 0.005619056988507509, + 0.003576153889298439, + 0.003316349582746625, + 0.004599760286509991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.50264972448349 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.406356930732727 + ], + "opposite_logprob_tokens": [ + -0.28135690093040466 + ], + "logprob_reward": -1.406356930732727, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406356930732727, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05631205067038536, + 0.04297720640897751, + 0.052236393094062805, + 0.04303134232759476, + 0.39971378445625305, + 0.09061285853385925, + 0.03846198692917824, + 0.025827990844845772, + 0.14560484886169434, + 0.10522153973579407 + ], + "credit_map": [ + 0.011271530762314796, + 0.0049960738979279995, + 0.007418854162096977, + 0.0039116512052714825, + 0.0041123004630208015, + 0.008390379138290882, + 0.0056047034449875355, + 0.010606561787426472, + 0.004766697995364666, + 0.005375794135034084, + 0.005596945062279701, + 0.007968117482960224, + 0.005693336017429829, + 0.0043715983629226685, + 0.003996079787611961, + 0.005208639893680811, + 0.005233316216617823, + 0.007299929857254028, + 0.006071678828448057, + 0.011878321878612041, + 0.004668937064707279, + 0.005336885340511799, + 0.005772374104708433, + 0.005974954459816217, + 0.00343084754422307, + 0.004664651583880186, + 0.0047080363146960735, + 0.005994055885821581, + 0.006385446526110172, + 0.006483619567006826, + 0.00441999826580286, + 0.006944691762328148, + 0.05417093262076378, + 0.08117404580116272, + 0.05531510338187218, + 0.050997380167245865, + 0.06189027428627014, + 0.03358561173081398, + 0.030653491616249084, + 0.0319269597530365, + 0.007581500336527824, + 0.023573672398924828, + 0.01763935200870037, + 0.009198839776217937, + 0.013742618262767792, + 0.0048477528616786, + 0.00738444784656167, + 0.006644683890044689, + 0.0023142206482589245, + 0.006325284484773874, + 0.004197751637548208, + 0.003975424915552139, + 0.011828397400677204, + 0.003616906004026532, + 0.0031027491204440594, + 0.0031012580730021, + 0.003445344977080822, + 0.0039088306948542595, + 0.002420401433482766, + 0.0029706242494285107, + 0.004613165743649006, + 0.002521324437111616, + 0.003038751892745495, + 0.0029095523059368134, + 0.013605930842459202, + 0.025719301775097847, + 0.026525532826781273, + 0.018746202811598778, + 0.0213331151753664, + 0.01238978374749422, + 0.011127923615276814, + 0.016157077625393867, + 0.013090252876281738, + 0.01563284918665886, + 0.012483733706176281, + 0.01334908977150917, + 0.014538075774908066, + 0.020266195759177208, + 0.005290459841489792, + 0.010570889338850975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -1.406356930732727 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014245601370930672 + ], + "opposite_logprob_tokens": [ + -4.264245510101318 + ], + "logprob_reward": -0.014245601370930672, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014245601370930672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13151943683624268, + 0.08543949574232101, + 0.09883753210306168, + 0.12730173766613007, + 0.16663435101509094, + 0.06496521830558777, + 0.03625595197081566, + 0.04077542945742607, + 0.09376493841409683, + 0.15450601279735565 + ], + "credit_map": [ + 0.028334660455584526, + 0.01467487495392561, + 0.01842227205634117, + 0.012513834051787853, + 0.012706143781542778, + 0.012498979456722736, + 0.012965683825314045, + 0.019402964040637016, + 0.012859262526035309, + 0.015338052995502949, + 0.011527917347848415, + 0.013672050088644028, + 0.008617989718914032, + 0.009058929048478603, + 0.007338132243603468, + 0.007027147337794304, + 0.011981491930782795, + 0.012012666091322899, + 0.013174168765544891, + 0.022539760917425156, + 0.008638888597488403, + 0.009596209041774273, + 0.010209094732999802, + 0.010685236193239689, + 0.009610670618712902, + 0.009913836605846882, + 0.016228679567575455, + 0.02697611227631569, + 0.017815185710787773, + 0.01736431010067463, + 0.014088393189013004, + 0.015304533764719963, + 0.022265303879976273, + 0.032499197870492935, + 0.018868358805775642, + 0.02326611429452896, + 0.022296493873000145, + 0.018702968955039978, + 0.013119579292833805, + 0.015616318210959435, + 0.0050062877126038074, + 0.01553349569439888, + 0.009068161249160767, + 0.008478703908622265, + 0.008824934251606464, + 0.004169878084212542, + 0.006806455086916685, + 0.007077289279550314, + 0.0026720818132162094, + 0.00817759707570076, + 0.0018055913969874382, + 0.004270221572369337, + 0.010276596993207932, + 0.0027688348200172186, + 0.0030592072289437056, + 0.00322581734508276, + 0.005075368098914623, + 0.0058909822255373, + 0.0050898389890789986, + 0.00481070764362812, + 0.0065648360177874565, + 0.0032542545814067125, + 0.0047288453206419945, + 0.0053605930879712105, + 0.009236445650458336, + 0.01824885606765747, + 0.0278936754912138, + 0.012723498977720737, + 0.008839694783091545, + 0.007948478683829308, + 0.004632466938346624, + 0.00424181530252099, + 0.01786779798567295, + 0.010820467956364155, + 0.033472463488578796, + 0.026901494711637497, + 0.014981876127421856, + 0.029637785628437996, + 0.007454482838511467, + 0.013369616121053696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.014245601370930672 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25216177105903625 + ], + "opposite_logprob_tokens": [ + -1.5021617412567139 + ], + "logprob_reward": -0.25216177105903625, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25216177105903625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0809241533279419, + 0.07224886119365692, + 0.11930259317159653, + 0.08970177173614502, + 0.06079364940524101, + 0.2048071324825287, + 0.06274615228176117, + 0.0377325713634491, + 0.07295972108840942, + 0.19878341257572174 + ], + "credit_map": [ + 0.011467044241726398, + 0.008472928777337074, + 0.008309929631650448, + 0.00802319310605526, + 0.007353332359343767, + 0.009802758693695068, + 0.010067895986139774, + 0.017427073791623116, + 0.006573662161827087, + 0.009527012705802917, + 0.008040694519877434, + 0.013979984447360039, + 0.008544213138520718, + 0.008299811743199825, + 0.010678667575120926, + 0.0066048153676092625, + 0.013959979638457298, + 0.015327776782214642, + 0.016911989077925682, + 0.03155440092086792, + 0.00791855901479721, + 0.009802701883018017, + 0.010859014466404915, + 0.012968169525265694, + 0.007200786843895912, + 0.008525696583092213, + 0.014748440124094486, + 0.01606200821697712, + 0.014792945235967636, + 0.010789462365210056, + 0.008141514845192432, + 0.009440911002457142, + 0.005584955681115389, + 0.008674235083162785, + 0.006219315342605114, + 0.00902036763727665, + 0.008387257345020771, + 0.008214036002755165, + 0.006501663941890001, + 0.008191813714802265, + 0.012508041225373745, + 0.06302635371685028, + 0.037705905735492706, + 0.017964622005820274, + 0.03752060979604721, + 0.008951717056334019, + 0.01471837516874075, + 0.012411518022418022, + 0.0038702378515154123, + 0.01111484132707119, + 0.01021089032292366, + 0.007396492641419172, + 0.010015388950705528, + 0.004898461513221264, + 0.00723263481631875, + 0.008007209748029709, + 0.004906323738396168, + 0.005304636433720589, + 0.0033175903372466564, + 0.004950591363012791, + 0.006009917240589857, + 0.0031993461307138205, + 0.004611899144947529, + 0.005432266276329756, + 0.006321195513010025, + 0.011964782141149044, + 0.018220042809844017, + 0.010391607880592346, + 0.009457334876060486, + 0.006130815949290991, + 0.004501286894083023, + 0.005972655490040779, + 0.02585959993302822, + 0.020530741661787033, + 0.029827184975147247, + 0.02677297033369541, + 0.02159205451607704, + 0.04258866235613823, + 0.010200007818639278, + 0.021412191912531853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.25216177105903625 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355251669883728 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.026355251669883728, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355251669883728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11409876495599747, + 0.10518743842840195, + 0.1195334792137146, + 0.03460075333714485, + 0.08693909645080566, + 0.18977680802345276, + 0.06037858501076698, + 0.11135037988424301, + 0.052652984857559204, + 0.12548166513442993 + ], + "credit_map": [ + 0.013503101654350758, + 0.012613591738045216, + 0.011044965125620365, + 0.011407321318984032, + 0.009701636619865894, + 0.014359069056808949, + 0.01476434525102377, + 0.026704730466008186, + 0.010419677011668682, + 0.011583669111132622, + 0.012699436396360397, + 0.01457307580858469, + 0.016445083543658257, + 0.01181150134652853, + 0.017700139433145523, + 0.009954852983355522, + 0.012481648474931717, + 0.021583357825875282, + 0.014128215610980988, + 0.02843518555164337, + 0.008589199744164944, + 0.008977574296295643, + 0.013364937156438828, + 0.011973361484706402, + 0.003882765071466565, + 0.004836086183786392, + 0.003946821670979261, + 0.006578836124390364, + 0.004945360589772463, + 0.004432779271155596, + 0.0027812609914690256, + 0.0031968418043106794, + 0.008206784725189209, + 0.009587129577994347, + 0.008333041332662106, + 0.011113115586340427, + 0.01109839603304863, + 0.016865763813257217, + 0.009700069203972816, + 0.012034803628921509, + 0.012161610648036003, + 0.055604904890060425, + 0.036500297486782074, + 0.01493602804839611, + 0.039050716906785965, + 0.009661421179771423, + 0.011729477904736996, + 0.010132352821528912, + 0.004804914817214012, + 0.007924886420369148, + 0.012629327364265919, + 0.00799905601888895, + 0.007200134918093681, + 0.005779892671853304, + 0.00626923143863678, + 0.007771142292767763, + 0.012143800035119057, + 0.018760312348604202, + 0.014243684709072113, + 0.013883210718631744, + 0.0176736768335104, + 0.008989211171865463, + 0.011937211267650127, + 0.013719271868467331, + 0.005117837339639664, + 0.013168670237064362, + 0.008026277646422386, + 0.00653252424672246, + 0.008776498027145863, + 0.0037420233711600304, + 0.0032550364267081022, + 0.004034116864204407, + 0.014406945556402206, + 0.009654083289206028, + 0.03424833342432976, + 0.02474167011678219, + 0.009676947258412838, + 0.018974045291543007, + 0.005020624492317438, + 0.00875900685787201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.026355251669883728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion with a static camera, the chain reaction is meticulously documented in a surreal, extreme close-up. A pristine, unmarked white aluminum aerosol can with a matte finish stands atop a dark, polished wood surface, its shape defined by intense, cool sidelighting, casting long shadows. The scene begins with low-frequency vibrations subtly vibrating the surface, causing the can to tremble almost imperceptibly. As the vibrations intensify, the can begins to tilt inexorably, its base slowly losing its grip on the smooth wood. It topples in a graceful arc, its white plastic actuator cap striking the hard edge of the wooden surface with a sharp, silent impact. The force of the collision instantly compresses the actuator, forcing a powerful, cone-shaped plume of fine mist from the nozzle. The mist billows outward, its individual particles catching the dramatic sidelight, forming a shimmering, ethereal cloud against the dark, out-of-focus background, showcasing the mechanics with clinical and stunning visual clarity.", + "seed": 7810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_6/index_1286_seed_7810_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": -0.3677974004298449, + "vlm_logprob": -0.3677974004298449, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.3677974004298449, + "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 video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39868220686912537 + ], + "opposite_logprob_tokens": [ + -1.3986822366714478 + ], + "logprob_reward": -0.39868220686912537, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.39868220686912537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025530286133289337, + 0.014192839153110981, + 0.04210721328854561, + 0.041820403188467026, + 0.11591628938913345, + 0.14429911971092224, + 0.12326696515083313, + 0.23100371658802032, + 0.21340210735797882, + 0.04846106097102165 + ], + "credit_map": [ + 0.0029967857990413904, + 0.0018984181806445122, + 0.0047634015791118145, + 0.004065505228936672, + 0.0019962096121162176, + 0.0027939684223383665, + 0.003801372367888689, + 0.0032146258745342493, + 0.0014751579146832228, + 0.0012396120000630617, + 0.0020704150665551424, + 0.002064973348751664, + 0.0009982646442949772, + 0.0024167154915630817, + 0.0015843830769881606, + 0.0023433174937963486, + 0.005844742059707642, + 0.003112186212092638, + 0.005765483248978853, + 0.005316801834851503, + 0.0026734829880297184, + 0.006707628257572651, + 0.0057472833432257175, + 0.006939604878425598, + 0.0039968532510101795, + 0.003585227532312274, + 0.007771113887429237, + 0.0058077494613826275, + 0.004719306714832783, + 0.005034774541854858, + 0.004001372493803501, + 0.006904005538672209, + 0.008890632539987564, + 0.012640975415706635, + 0.011929076164960861, + 0.016622086986899376, + 0.013370624743402004, + 0.018646540120244026, + 0.014786826446652412, + 0.01902952417731285, + 0.008319804444909096, + 0.008228237740695477, + 0.008451172150671482, + 0.005416591186076403, + 0.061215825378894806, + 0.01189179066568613, + 0.00502970814704895, + 0.03574599698185921, + 0.006522243842482567, + 0.008847150020301342, + 0.009024093858897686, + 0.00926108006387949, + 0.02187160961329937, + 0.035306282341480255, + 0.008030697703361511, + 0.02440381981432438, + 0.01020400132983923, + 0.008459962904453278, + 0.013201895169913769, + 0.011982744559645653, + 0.02028379589319229, + 0.08547544479370117, + 0.009124425239861012, + 0.0722714513540268, + 0.01146072056144476, + 0.012154921889305115, + 0.008199485950171947, + 0.0042397938668727875, + 0.08182530105113983, + 0.07912140339612961, + 0.0053227622993290424, + 0.011077729985117912, + 0.005673584062606096, + 0.0015061565209180117, + 0.007163753267377615, + 0.003688910510390997, + 0.003744164016097784, + 0.005122775677591562, + 0.005501930136233568, + 0.016059787943959236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.39868220686912537 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.3134815692901611 + ], + "opposite_logprob_tokens": [ + -0.3134815990924835 + ], + "logprob_reward": -1.3134815692901611, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134815692901611, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05322391167283058, + 0.054148346185684204, + 0.029425112530589104, + 0.04554139822721481, + 0.05105643346905708, + 0.0823923870921135, + 0.09935946762561798, + 0.4136348366737366, + 0.13473248481750488, + 0.03648560121655464 + ], + "credit_map": [ + 0.005351541563868523, + 0.003485301276668906, + 0.009191195480525494, + 0.00936818029731512, + 0.00846859160810709, + 0.0050815106369555, + 0.007547830697149038, + 0.004729758016765118, + 0.0033397164661437273, + 0.001375603605993092, + 0.017263950780034065, + 0.0073876758106052876, + 0.003611278021708131, + 0.01439350750297308, + 0.003930944483727217, + 0.0028456710278987885, + 0.004644427448511124, + 0.001814288436435163, + 0.0050328681245446205, + 0.0035573896020650864, + 0.0018625635420903563, + 0.004744968842715025, + 0.003557379823178053, + 0.004211226012557745, + 0.005399826914072037, + 0.002749058650806546, + 0.010886024683713913, + 0.005215030629187822, + 0.005719025619328022, + 0.004432375077158213, + 0.003369829384610057, + 0.007770226802676916, + 0.0030021246057003736, + 0.0033706314861774445, + 0.003925977274775505, + 0.004230962134897709, + 0.0050892820581793785, + 0.01081879623234272, + 0.00592662813141942, + 0.014692028053104877, + 0.004394332878291607, + 0.004969018045812845, + 0.005159780848771334, + 0.003446882124990225, + 0.03465208411216736, + 0.00731630576774478, + 0.0037501107435673475, + 0.018703877925872803, + 0.005050857085734606, + 0.007374713197350502, + 0.007219433318823576, + 0.007824837230145931, + 0.019395317882299423, + 0.02719978243112564, + 0.00798916257917881, + 0.017305362969636917, + 0.019463608041405678, + 0.014976718463003635, + 0.026594460010528564, + 0.024543438106775284, + 0.03302454948425293, + 0.15007704496383667, + 0.019408434629440308, + 0.12554657459259033, + 0.013143849559128284, + 0.0067751542665064335, + 0.010197736322879791, + 0.005322874523699284, + 0.043249912559986115, + 0.011991811916232109, + 0.007254761643707752, + 0.03679637983441353, + 0.005930067040026188, + 0.0011184829054400325, + 0.011030243709683418, + 0.0030511757358908653, + 0.002210112987086177, + 0.006963244639337063, + 0.0027543255127966404, + 0.003427951131016016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -1.3134815692901611 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016095465049147606 + ], + "opposite_logprob_tokens": [ + -4.1410956382751465 + ], + "logprob_reward": -0.016095465049147606, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016095465049147606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029493045061826706, + 0.028509892523288727, + 0.02364908531308174, + 0.04580363631248474, + 0.11477968096733093, + 0.060788799077272415, + 0.05762705206871033, + 0.3218456208705902, + 0.2990381419658661, + 0.018465036526322365 + ], + "credit_map": [ + 0.0026245173066854477, + 0.0023196230176836252, + 0.004361475817859173, + 0.004564187955111265, + 0.005425155162811279, + 0.003195809666067362, + 0.004143217112869024, + 0.002859057392925024, + 0.002156101865693927, + 0.0012564759235829115, + 0.007231794763356447, + 0.003385016694664955, + 0.002407695399597287, + 0.007979529909789562, + 0.0020047430880367756, + 0.0020885332487523556, + 0.0035765771754086018, + 0.001833396963775158, + 0.002888821065425873, + 0.003022625809535384, + 0.001402512425556779, + 0.004159724339842796, + 0.002073955489322543, + 0.004691472742706537, + 0.004745219834148884, + 0.002847307361662388, + 0.006956176832318306, + 0.005782251711934805, + 0.005733359605073929, + 0.01035951729863882, + 0.005128353834152222, + 0.004251450765877962, + 0.007805324625223875, + 0.009400204755365849, + 0.008835542015731335, + 0.012611190788447857, + 0.00918665062636137, + 0.02715860679745674, + 0.011751765385270119, + 0.0280303917825222, + 0.004152074921876192, + 0.004176144488155842, + 0.0044752382673323154, + 0.003397817723453045, + 0.024568727239966393, + 0.009473740123212337, + 0.003275069873780012, + 0.007269990164786577, + 0.003750127274543047, + 0.00436322670429945, + 0.007306460291147232, + 0.005767991300672293, + 0.012912631966173649, + 0.011242757551372051, + 0.00657255295664072, + 0.005711305886507034, + 0.01576627977192402, + 0.011640669777989388, + 0.01846257969737053, + 0.018057754263281822, + 0.019972866401076317, + 0.11557146906852722, + 0.015418753027915955, + 0.10695524513721466, + 0.015438207425177097, + 0.016562247648835182, + 0.012518207542598248, + 0.005068541504442692, + 0.11221387982368469, + 0.1104493960738182, + 0.00734991068020463, + 0.01943773217499256, + 0.0015953759429976344, + 0.0010371776297688484, + 0.0019042434869334102, + 0.0017299067694693804, + 0.005128349643200636, + 0.002373324939981103, + 0.0015938144642859697, + 0.003102844348177314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.016095465049147606 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048683732748031616 + ], + "opposite_logprob_tokens": [ + -3.0486836433410645 + ], + "logprob_reward": -0.048683732748031616, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683732748031616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010999192483723164, + 0.03365340456366539, + 0.011010776273906231, + 0.014748902060091496, + 0.02183246798813343, + 0.03739041090011597, + 0.01600310020148754, + 0.42153722047805786, + 0.386443167924881, + 0.04638133570551872 + ], + "credit_map": [ + 0.0012987495865672827, + 0.0010000169277191162, + 0.0017097210511565208, + 0.00184021913446486, + 0.0009423909359611571, + 0.0013118452625349164, + 0.001602863660082221, + 0.0012933859834447503, + 0.001980871893465519, + 0.0006569697870872915, + 0.011845936067402363, + 0.00478864973410964, + 0.0016879575559869409, + 0.008786999620497227, + 0.002433918882161379, + 0.0014721029438078403, + 0.001588163198903203, + 0.001019166549667716, + 0.0013425200013443828, + 0.001493163057602942, + 0.0007871694397181273, + 0.001926449011079967, + 0.0015276351477950811, + 0.0013265105662867427, + 0.0019438114250078797, + 0.0014975148951634765, + 0.0019078570185229182, + 0.0019030262483283877, + 0.0018857468385249376, + 0.002052069641649723, + 0.0017570428317412734, + 0.001801833393983543, + 0.0007949094870127738, + 0.0011404104297980666, + 0.0011766336392611265, + 0.00141714618075639, + 0.0013710015919059515, + 0.0067736865021288395, + 0.0011671934043988585, + 0.007991488091647625, + 0.002023365581408143, + 0.0020066683646291494, + 0.0024023340083658695, + 0.001392663107253611, + 0.015614555217325687, + 0.0029892525635659695, + 0.0015262465458363295, + 0.009435324929654598, + 0.000747174839489162, + 0.0007584613049402833, + 0.0010245998855680227, + 0.0010306538315489888, + 0.000923933694139123, + 0.005115089938044548, + 0.0008057186496444046, + 0.005597467999905348, + 0.018668225035071373, + 0.015407061204314232, + 0.024652203544974327, + 0.022533312439918518, + 0.03665683791041374, + 0.15582647919654846, + 0.016109997406601906, + 0.1316831260919571, + 0.02102532796561718, + 0.02361896261572838, + 0.015408186241984367, + 0.006451067980378866, + 0.15952205657958984, + 0.13905860483646393, + 0.010989787057042122, + 0.010369179770350456, + 0.0026466716080904007, + 0.0013621572870761156, + 0.005338391754776239, + 0.0017080854158848524, + 0.01465093158185482, + 0.01549388375133276, + 0.001742320484481752, + 0.00343889556825161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.048683732748031616 + }, + { + "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 shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the shot an extreme close-up of a white aluminum aerosol can?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the can tilt and topple over due to vibrations before striking the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the can release a cone-shaped plume of mist upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204402819275856 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.06204402819275856, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204402819275856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03521176800131798, + 0.04451260343194008, + 0.04861684888601303, + 0.2055417001247406, + 0.04940708354115486, + 0.030696365982294083, + 0.02387990429997444, + 0.08311467617750168, + 0.30474647879600525, + 0.17427262663841248 + ], + "credit_map": [ + 0.00476772291585803, + 0.0028659990057349205, + 0.006185552570968866, + 0.004887427669018507, + 0.0030489624477922916, + 0.004632372874766588, + 0.004698224365711212, + 0.0041255047544837, + 0.0027488581836223602, + 0.0010606020223349333, + 0.016157817095518112, + 0.006856818217784166, + 0.001586299971677363, + 0.0104287164285779, + 0.0036723287776112556, + 0.0020011593587696552, + 0.008010899648070335, + 0.003538097022101283, + 0.008689997717738152, + 0.007745333015918732, + 0.0029061383102089167, + 0.00475237937644124, + 0.007803885731846094, + 0.00517011433839798, + 0.026752492412924767, + 0.013496615923941135, + 0.04466009512543678, + 0.023221787065267563, + 0.01835457794368267, + 0.0367795005440712, + 0.01803150586783886, + 0.024245094507932663, + 0.0028933347202837467, + 0.0034465654753148556, + 0.003392208367586136, + 0.004254128318279982, + 0.0044761584140360355, + 0.012542531825602055, + 0.003388711716979742, + 0.015013438649475574, + 0.0013468698598444462, + 0.0019385638879612088, + 0.0017533493228256702, + 0.0015538407023996115, + 0.009192257188260555, + 0.003300474723801017, + 0.0018483835738152266, + 0.009762623347342014, + 0.0011371176224201918, + 0.0018188919639214873, + 0.001660950481891632, + 0.0016561556840315461, + 0.005910131148993969, + 0.006903492379933596, + 0.001959591405466199, + 0.002833571517840028, + 0.0033156012650579214, + 0.0037884199991822243, + 0.004308982752263546, + 0.004500961396843195, + 0.010960250161588192, + 0.030861128121614456, + 0.003221978200599551, + 0.022157352417707443, + 0.0169393178075552, + 0.0135576156899333, + 0.013074464164674282, + 0.006980126723647118, + 0.09194088727235794, + 0.11589409410953522, + 0.0073952036909759045, + 0.03896475210785866, + 0.015399765223264694, + 0.0050394549034535885, + 0.03167952969670296, + 0.012500083073973656, + 0.026944514364004135, + 0.05778465420007706, + 0.010620874352753162, + 0.014303735457360744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.06204402819275856 + } + ], + "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/0000017/group_7/index_tensor([8287, 8287])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_tensor([8287, 8287])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..310409fae2db563ffcedf1c645b3adbc15e16d54 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_tensor([8287, 8287])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from a level angle, slowly and steadily pulls back, gradually revealing a vast bamboo forest shrouded in thick morning mist. At the center of the frame, an elderly man with a youthful face, his long, silvery hair neatly tied into a Taoist bun, clad in a loose-fitting, pure white silk training suit embellished with traditional handmade buttons. He is immersed in a Tai Chi practice, his arms tracing smooth, flowing arcs, his body turning smoothly with the shifting of his center of gravity. Each movement is imbued with inner strength and unhurried rhythm. Soft morning light penetrates the mist and the tall, emerald green bamboo stalks, casting distinct beams of light in the air, casting a subtle shift of light and shadow upon the elderly man and the leaf-strewn ground around him. The scene's dominant color palette is the cool emerald green of the bamboo forest and the milky white of the mist, creating a tranquil, ethereal, and Zen-like atmosphere imbued with Eastern aesthetics.", + "seed": 15154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15154_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": -0.05602331701666117, + "vlm_logprob": -0.05602331701666117, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05602331701666117, + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.019697068259119987 + ], + "opposite_logprob_tokens": [ + -4.144697189331055 + ], + "logprob_reward": -0.019697068259119987, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019697068259119987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14386864006519318, + 0.12486586719751358, + 0.12139579653739929, + 0.07195387780666351, + 0.08672041445970535, + 0.09761395305395126, + 0.08209509402513504, + 0.11068572849035263, + 0.13082069158554077, + 0.029979916289448738 + ], + "credit_map": [ + 0.025212811306118965, + 0.020756114274263382, + 0.01688266731798649, + 0.02072415128350258, + 0.01788773015141487, + 0.012240681797266006, + 0.020580917596817017, + 0.009583563543856144, + 0.012018004432320595, + 0.017052283510565758, + 0.010267844423651695, + 0.015863416716456413, + 0.015930304303765297, + 0.02356121502816677, + 0.0159035325050354, + 0.01426927000284195, + 0.013220462016761303, + 0.015855105593800545, + 0.01064182911068201, + 0.020502056926488876, + 0.025968970730900764, + 0.016000637784600258, + 0.010077179409563541, + 0.00912956427782774, + 0.00484688114374876, + 0.009206999093294144, + 0.00512732844799757, + 0.011648072861135006, + 0.015627063810825348, + 0.005146894138306379, + 0.013825696893036366, + 0.006524941883981228, + 0.0068822698667645454, + 0.009077200666069984, + 0.006933002267032862, + 0.007773444056510925, + 0.020996414124965668, + 0.010745572857558727, + 0.014909790828824043, + 0.00940271932631731, + 0.012109608389437199, + 0.01654311828315258, + 0.0114668607711792, + 0.009084978140890598, + 0.012594413012266159, + 0.0183434896171093, + 0.008053682744503021, + 0.009417803026735783, + 0.008086438290774822, + 0.015881726518273354, + 0.010813434608280659, + 0.011916624382138252, + 0.012437015771865845, + 0.00829230435192585, + 0.006002049893140793, + 0.00866550114005804, + 0.012673325836658478, + 0.01563543826341629, + 0.011726743541657925, + 0.008793890476226807, + 0.017195435240864754, + 0.019908830523490906, + 0.01076078787446022, + 0.013991283252835274, + 0.021730374544858932, + 0.012493658810853958, + 0.020432408899068832, + 0.008370138704776764, + 0.01479672733694315, + 0.023523105308413506, + 0.012064418755471706, + 0.017409848049283028, + 0.004729366395622492, + 0.0036492599174380302, + 0.003315010340884328, + 0.002860091160982847, + 0.004702923819422722, + 0.003914075903594494, + 0.0030555601697415113, + 0.0037536281161010265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.019697068259119987 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.014218217693269253 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.014218217693269253, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218217693269253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10829473286867142, + 0.11059186607599258, + 0.31109359860420227, + 0.11555499583482742, + 0.0691632553935051, + 0.05204043164849281, + 0.07862025499343872, + 0.06593307107686996, + 0.03702127933502197, + 0.05168648436665535 + ], + "credit_map": [ + 0.017048470675945282, + 0.01786026731133461, + 0.011938572861254215, + 0.012197544798254967, + 0.017458634451031685, + 0.009574398398399353, + 0.011553852818906307, + 0.010662992484867573, + 0.010139909572899342, + 0.01634826511144638, + 0.010680979117751122, + 0.015491562895476818, + 0.01531505212187767, + 0.01907655969262123, + 0.011838293634355068, + 0.011701243929564953, + 0.10621850192546844, + 0.03148543834686279, + 0.030840732157230377, + 0.013530063442885876, + 0.016468055546283722, + 0.033192191272974014, + 0.012574009597301483, + 0.06678460538387299, + 0.00816228985786438, + 0.022224778309464455, + 0.010076433420181274, + 0.012812518514692783, + 0.02216651476919651, + 0.010707610286772251, + 0.017304882407188416, + 0.012099958024919033, + 0.0077794743701815605, + 0.006757314782589674, + 0.008375233970582485, + 0.006394608877599239, + 0.008905666880309582, + 0.015644973143935204, + 0.00958421640098095, + 0.005721768829971552, + 0.007087172009050846, + 0.007608485873788595, + 0.006664900109171867, + 0.005056602414697409, + 0.004424686543643475, + 0.011909441091120243, + 0.004817857872694731, + 0.004471283871680498, + 0.0074148597195744514, + 0.013022957369685173, + 0.011457542888820171, + 0.0072229402139782906, + 0.013029846362769604, + 0.011480048298835754, + 0.006356484722346067, + 0.008635575883090496, + 0.008512590080499649, + 0.00881039910018444, + 0.012662438675761223, + 0.006685825064778328, + 0.010029706172645092, + 0.0061335209757089615, + 0.006186446640640497, + 0.006912148557603359, + 0.0047746081836521626, + 0.0037952098064124584, + 0.00768304243683815, + 0.003959472756832838, + 0.0044890339486300945, + 0.003958896733820438, + 0.0038987325970083475, + 0.0044622840359807014, + 0.00939610879868269, + 0.006295122671872377, + 0.005682811141014099, + 0.003596983850002289, + 0.008397729136049747, + 0.0062317312695086, + 0.006338763050734997, + 0.005747235845774412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.014218217693269253 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.043062031269073486 + ], + "opposite_logprob_tokens": [ + -3.1680619716644287 + ], + "logprob_reward": -0.043062031269073486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043062031269073486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05444923788309097, + 0.030754294246435165, + 0.7150353193283081, + 0.02144523523747921, + 0.048848867416381836, + 0.017399301752448082, + 0.029109131544828415, + 0.023187998682260513, + 0.025192048400640488, + 0.034578587859869 + ], + "credit_map": [ + 0.005185998976230621, + 0.01008221972733736, + 0.004779840353876352, + 0.009335911832749844, + 0.007583776954561472, + 0.0057402984239161015, + 0.006902808323502541, + 0.004838381893932819, + 0.0036099920980632305, + 0.004375779069960117, + 0.002974848262965679, + 0.0043241665698587894, + 0.005371704697608948, + 0.003172418335452676, + 0.0035201851278543472, + 0.00340519892051816, + 0.23625022172927856, + 0.07089898735284805, + 0.0673900917172432, + 0.038739562034606934, + 0.04136734828352928, + 0.07666639238595963, + 0.03136912360787392, + 0.15235359966754913, + 0.0019171311287209392, + 0.003246017498895526, + 0.0027200975455343723, + 0.002866805763915181, + 0.003249926958233118, + 0.0025859663728624582, + 0.0028073994908481836, + 0.002051890827715397, + 0.004211696796119213, + 0.005578103009611368, + 0.007471338380128145, + 0.00623714504763484, + 0.007472192402929068, + 0.005747527349740267, + 0.007944413460791111, + 0.0041864472441375256, + 0.002287813927978277, + 0.0025471439585089684, + 0.0018666577525436878, + 0.0018158042803406715, + 0.0019048515241593122, + 0.003591893706470728, + 0.001711013144813478, + 0.0016741234576329589, + 0.0030728639103472233, + 0.004866014234721661, + 0.0028225723654031754, + 0.002571498742327094, + 0.0025208787992596626, + 0.005877544637769461, + 0.0025809400249272585, + 0.004796819761395454, + 0.0036453481297940016, + 0.002298512728884816, + 0.0022072040010243654, + 0.0016688231844455004, + 0.004864422604441643, + 0.003224252490326762, + 0.0026374340523034334, + 0.0026420021895319223, + 0.0032653105445206165, + 0.00194651132915169, + 0.004952652379870415, + 0.002353732008486986, + 0.004186664242297411, + 0.0030819722451269627, + 0.0033342381939291954, + 0.0020709678065031767, + 0.005592855159193277, + 0.00518085015937686, + 0.004285519942641258, + 0.002695455215871334, + 0.004706129431724548, + 0.003849072614684701, + 0.004134065471589565, + 0.004134639631956816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.043062031269073486 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09613320231437683, + 0.1755146086215973, + 0.1928902119398117, + 0.03762420639395714, + 0.08833746612071991, + 0.10543078929185867, + 0.12472992390394211, + 0.13566835224628448, + 0.015986710786819458, + 0.027684470638632774 + ], + "credit_map": [ + 0.013027172535657883, + 0.008800654672086239, + 0.016173068434000015, + 0.010437889024615288, + 0.02052091620862484, + 0.01116681657731533, + 0.009706290438771248, + 0.006300398614257574, + 0.022993985563516617, + 0.011932946741580963, + 0.012731075286865234, + 0.014688227325677872, + 0.024401873350143433, + 0.04785183444619179, + 0.012010729871690273, + 0.028903944417834282, + 0.06651414930820465, + 0.019256308674812317, + 0.017400728538632393, + 0.009974416345357895, + 0.009200871922075748, + 0.019976025447249413, + 0.008992047049105167, + 0.04157567396759987, + 0.0038668096531182528, + 0.004541479982435703, + 0.00319630466401577, + 0.006441876292228699, + 0.005960410926491022, + 0.003942603711038828, + 0.0067181834019720554, + 0.002956541022285819, + 0.013855736702680588, + 0.011516758240759373, + 0.017277229577302933, + 0.009290170855820179, + 0.005728256423026323, + 0.01673426851630211, + 0.008867009542882442, + 0.005068046506494284, + 0.014777020551264286, + 0.016651717945933342, + 0.01737646386027336, + 0.011712142266333103, + 0.006338819861412048, + 0.023723863065242767, + 0.008202972821891308, + 0.006647797301411629, + 0.018896596506237984, + 0.02291274629533291, + 0.022907817736268044, + 0.013137311674654484, + 0.012285668402910233, + 0.016598032787442207, + 0.007513679098337889, + 0.010478082112967968, + 0.025302637368440628, + 0.017627820372581482, + 0.026185140013694763, + 0.01410703081637621, + 0.007376319263130426, + 0.026488013565540314, + 0.007568880449980497, + 0.011012517847120762, + 0.0021306946873664856, + 0.002239092718809843, + 0.002537699416279793, + 0.0011348447296768427, + 0.0025245666038244963, + 0.0020400332286953926, + 0.0017427513375878334, + 0.001637027133256197, + 0.004549803677946329, + 0.003632906824350357, + 0.005158102139830589, + 0.002571140881627798, + 0.0031715775839984417, + 0.0029810923151671886, + 0.003211800241842866, + 0.0024080504663288593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.023279689252376556 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.17985957860946655 + ], + "opposite_logprob_tokens": [ + -1.8048596382141113 + ], + "logprob_reward": -0.17985957860946655, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985957860946655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14946670830249786, + 0.08037763833999634, + 0.386252760887146, + 0.04423951730132103, + 0.04484555870294571, + 0.03669135645031929, + 0.06630752235651016, + 0.06311019510030746, + 0.060782790184020996, + 0.06792595237493515 + ], + "credit_map": [ + 0.016424510627985, + 0.025277500972151756, + 0.016736231744289398, + 0.01271851547062397, + 0.02524336613714695, + 0.015942972153425217, + 0.018219083547592163, + 0.018904531374573708, + 0.008519583381712437, + 0.010665599256753922, + 0.0060926093719899654, + 0.009407381527125835, + 0.010509097017347813, + 0.01612255722284317, + 0.008471841923892498, + 0.010588975623250008, + 0.12433968484401703, + 0.0386928915977478, + 0.0336947962641716, + 0.023169906809926033, + 0.025808390229940414, + 0.04113832488656044, + 0.017637677490711212, + 0.081771120429039, + 0.003567717969417572, + 0.0083119822666049, + 0.0031866314820945263, + 0.006607875227928162, + 0.007066505961120129, + 0.0035829355474561453, + 0.007223850581794977, + 0.00469202408567071, + 0.004328631795942783, + 0.008190439082682133, + 0.005664018448442221, + 0.005426050629466772, + 0.006005469709634781, + 0.0036928888875991106, + 0.007678233087062836, + 0.003859830554574728, + 0.0053403484635055065, + 0.005824976600706577, + 0.005050305277109146, + 0.0046567716635763645, + 0.0023388611152768135, + 0.008867477998137474, + 0.0028668262530118227, + 0.001745793386362493, + 0.009673970751464367, + 0.012111557647585869, + 0.010313322767615318, + 0.00863354280591011, + 0.00588188786059618, + 0.00984957441687584, + 0.004234036430716515, + 0.005609630607068539, + 0.013531391508877277, + 0.006601609289646149, + 0.005412745755165815, + 0.004384397529065609, + 0.005834598559886217, + 0.017123503610491753, + 0.005132327321916819, + 0.0050896285101771355, + 0.007987153716385365, + 0.006301357876509428, + 0.008739239536225796, + 0.004420023411512375, + 0.0071357483975589275, + 0.007976439781486988, + 0.005991724319756031, + 0.012231106869876385, + 0.010555729269981384, + 0.012934420257806778, + 0.011686259880661964, + 0.005629643797874451, + 0.009589974768459797, + 0.0062772901728749275, + 0.007132428698241711, + 0.00412021204829216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.17985957860946655 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a level angle, slowly and steadily pulls back, gradually revealing a vast bamboo forest shrouded in thick morning mist. At the center of the frame, an elderly man with a youthful face, his long, silvery hair neatly tied into a Taoist bun, clad in a loose-fitting, pure white silk training suit embellished with traditional handmade buttons. He is immersed in a Tai Chi practice, his arms tracing smooth, flowing arcs, his body turning smoothly with the shifting of his center of gravity. Each movement is imbued with inner strength and unhurried rhythm. Soft morning light penetrates the mist and the tall, emerald green bamboo stalks, casting distinct beams of light in the air, casting a subtle shift of light and shadow upon the elderly man and the leaf-strewn ground around him. The scene's dominant color palette is the cool emerald green of the bamboo forest and the milky white of the mist, creating a tranquil, ethereal, and Zen-like atmosphere imbued with Eastern aesthetics.", + "seed": 15155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15155_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": -0.027260722406208516, + "vlm_logprob": -0.027260722406208516, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.027260722406208516, + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.01783149316906929 + ], + "opposite_logprob_tokens": [ + -4.517831325531006 + ], + "logprob_reward": -0.01783149316906929, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01783149316906929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07655062526464462, + 0.2580290138721466, + 0.06335769593715668, + 0.07728203386068344, + 0.018374120816588402, + 0.036306433379650116, + 0.07170479744672775, + 0.038819387555122375, + 0.32738369703292847, + 0.03219223767518997 + ], + "credit_map": [ + 0.005984493996948004, + 0.005731887184083462, + 0.008360505104064941, + 0.009947356767952442, + 0.01607973501086235, + 0.004959370009601116, + 0.013731773011386395, + 0.01175550278276205, + 0.01074132602661848, + 0.02031179703772068, + 0.02277444675564766, + 0.10457752645015717, + 0.013469619676470757, + 0.006937389262020588, + 0.01664937287569046, + 0.06256751716136932, + 0.005430669989436865, + 0.0054702977649867535, + 0.006571268197149038, + 0.009695785120129585, + 0.008827782236039639, + 0.004420273005962372, + 0.015323685482144356, + 0.007617930881679058, + 0.012115581892430782, + 0.008197966031730175, + 0.009239019826054573, + 0.007710156496614218, + 0.007462042849510908, + 0.015493149869143963, + 0.010096819140017033, + 0.006967297755181789, + 0.00222677830606699, + 0.003028943669050932, + 0.0028136810287833214, + 0.0018038238631561399, + 0.0033262292854487896, + 0.001989196753129363, + 0.0015094216214492917, + 0.0016760462895035744, + 0.003682984970510006, + 0.003874122863635421, + 0.004470326472073793, + 0.00284213712438941, + 0.0034597045741975307, + 0.006166200153529644, + 0.00309914443641901, + 0.008711813017725945, + 0.007703241892158985, + 0.00874991063028574, + 0.007576781325042248, + 0.008444762788712978, + 0.010516736656427383, + 0.008637255989015102, + 0.01082927081733942, + 0.009246839210391045, + 0.003579305950552225, + 0.003693183185532689, + 0.0039314026944339275, + 0.00355564896017313, + 0.002899890299886465, + 0.007607733365148306, + 0.0030707120895385742, + 0.010481509380042553, + 0.026335589587688446, + 0.008285039104521275, + 0.017158908769488335, + 0.010048143565654755, + 0.018825294449925423, + 0.11732883751392365, + 0.012008043006062508, + 0.11739382892847061, + 0.005518189165741205, + 0.002786763943731785, + 0.005329297389835119, + 0.002511660335585475, + 0.0035540536046028137, + 0.004511173814535141, + 0.0033382046967744827, + 0.004642895422875881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.01783149316906929 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.012573259882628918 + ], + "opposite_logprob_tokens": [ + -4.3875732421875 + ], + "logprob_reward": -0.012573259882628918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012573259882628918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10557504743337631, + 0.06675232946872711, + 0.04360702633857727, + 0.0476299449801445, + 0.02660214900970459, + 0.047588229179382324, + 0.061570510268211365, + 0.03448678180575371, + 0.5280445218086243, + 0.03814353793859482 + ], + "credit_map": [ + 0.01252597663551569, + 0.012317754328250885, + 0.013454566709697247, + 0.007780785206705332, + 0.018758321180939674, + 0.008089136332273483, + 0.01347779855132103, + 0.019170690327882767, + 0.0036100258585065603, + 0.005192360375076532, + 0.006507676560431719, + 0.02669387124478817, + 0.004317553248256445, + 0.0018458683043718338, + 0.004960827995091677, + 0.013624140061438084, + 0.004700069315731525, + 0.005769859999418259, + 0.007929056882858276, + 0.005340928211808205, + 0.006120968144387007, + 0.0038803289644420147, + 0.006109389942139387, + 0.00375641998834908, + 0.007958847098052502, + 0.004338452592492104, + 0.006238920614123344, + 0.0037331553176045418, + 0.0052849305793643, + 0.009543134830892086, + 0.0065234447829425335, + 0.0040090493857860565, + 0.003509429981932044, + 0.0035508861765265465, + 0.0031229588203132153, + 0.004116362892091274, + 0.0028232731856405735, + 0.002740598749369383, + 0.0033339890651404858, + 0.0034046475775539875, + 0.005750053562223911, + 0.005991814192384481, + 0.0075761196203529835, + 0.004860034212470055, + 0.004811620805412531, + 0.0066272299736738205, + 0.004243839532136917, + 0.007727507501840591, + 0.006397681310772896, + 0.00656687468290329, + 0.006921905092895031, + 0.0061815958470106125, + 0.008281053975224495, + 0.0069491006433963776, + 0.009317981079220772, + 0.010954310186207294, + 0.003308128798380494, + 0.003730481956154108, + 0.005266003310680389, + 0.0031329544726759195, + 0.004353106487542391, + 0.004495786502957344, + 0.005169621668756008, + 0.005030692555010319, + 0.04339973255991936, + 0.01945183053612709, + 0.029670439660549164, + 0.0199898649007082, + 0.029489904642105103, + 0.18288257718086243, + 0.020898154005408287, + 0.18226198852062225, + 0.0040452741086483, + 0.005035048350691795, + 0.004936083219945431, + 0.004191952757537365, + 0.006923547480255365, + 0.0032123897690325975, + 0.004111347254365683, + 0.0056878868490457535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.012573259882628918 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.023292500525712967 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.023292500525712967, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292500525712967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09183372557163239, + 0.15317361056804657, + 0.03215496242046356, + 0.1036696508526802, + 0.025142934173345566, + 0.05229390412569046, + 0.10118422657251358, + 0.12745410203933716, + 0.25535961985588074, + 0.05773323401808739 + ], + "credit_map": [ + 0.009343472309410572, + 0.009986335411667824, + 0.014335669577121735, + 0.009566707536578178, + 0.014496858231723309, + 0.0072819241322577, + 0.011572668328881264, + 0.015250093303620815, + 0.006901522167026997, + 0.011797984130680561, + 0.013165087439119816, + 0.06243857368826866, + 0.008011861704289913, + 0.004373867064714432, + 0.009543030522763729, + 0.036941688507795334, + 0.0027719284407794476, + 0.004268701188266277, + 0.005792652256786823, + 0.004710929002612829, + 0.0032171898055821657, + 0.0027958571445196867, + 0.005723188165575266, + 0.0028745143208652735, + 0.017324430868029594, + 0.012055809609591961, + 0.012278079986572266, + 0.011478795669972897, + 0.010770967230200768, + 0.019539164379239082, + 0.010021197609603405, + 0.010201207362115383, + 0.003032105043530464, + 0.0030314556788653135, + 0.0024088381323963404, + 0.0033217272721230984, + 0.004304582253098488, + 0.0027704485692083836, + 0.0030438893008977175, + 0.0032298872247338295, + 0.004964057356119156, + 0.006504144053906202, + 0.006563007365912199, + 0.008649208582937717, + 0.006244239863008261, + 0.006333296187222004, + 0.0059297895058989525, + 0.007106164004653692, + 0.011471046134829521, + 0.012277874164283276, + 0.009603982791304588, + 0.011236361227929592, + 0.012811942957341671, + 0.011313186027109623, + 0.015858203172683716, + 0.016611630097031593, + 0.010119347833096981, + 0.017230456694960594, + 0.013540174812078476, + 0.012241367250680923, + 0.015765219926834106, + 0.015486951917409897, + 0.01828969642519951, + 0.02478089928627014, + 0.022663120180368423, + 0.027876274660229683, + 0.024775108322501183, + 0.019817685708403587, + 0.019550688564777374, + 0.0628000870347023, + 0.018367407843470573, + 0.05950925499200821, + 0.011419638060033321, + 0.00587109150364995, + 0.007711722049862146, + 0.00420029554516077, + 0.0050962744280695915, + 0.009032176807522774, + 0.005823919549584389, + 0.008578113280236721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.023292500525712967 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.020578688010573387 + ], + "opposite_logprob_tokens": [ + -3.895578622817993 + ], + "logprob_reward": -0.020578688010573387, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578688010573387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1521940678358078, + 0.18046684563159943, + 0.038779933005571365, + 0.11551738530397415, + 0.06780339032411575, + 0.14553691446781158, + 0.06257841736078262, + 0.05733124911785126, + 0.06999345868825912, + 0.10979829728603363 + ], + "credit_map": [ + 0.015426977537572384, + 0.014291219413280487, + 0.019348815083503723, + 0.017219338566064835, + 0.02495003119111061, + 0.011736851185560226, + 0.020933352410793304, + 0.028287500143051147, + 0.011875462718307972, + 0.01522529125213623, + 0.021048085764050484, + 0.07224325835704803, + 0.007212875410914421, + 0.005000526085495949, + 0.014595146290957928, + 0.03326622024178505, + 0.00408795615658164, + 0.004055465571582317, + 0.004821784794330597, + 0.008842346258461475, + 0.003332608612254262, + 0.004911607131361961, + 0.004977311473339796, + 0.0037508527748286724, + 0.019862553104758263, + 0.010747983120381832, + 0.015670394524931908, + 0.010777268558740616, + 0.010213390924036503, + 0.022511925548315048, + 0.015568673610687256, + 0.0101652005687356, + 0.010256859473884106, + 0.010341771878302097, + 0.010328926146030426, + 0.007335026748478413, + 0.007668428122997284, + 0.007451048586517572, + 0.007162411697208881, + 0.00725892186164856, + 0.0192656721919775, + 0.01763402298092842, + 0.02253562957048416, + 0.0155626917257905, + 0.012491006404161453, + 0.02143438719213009, + 0.010726168751716614, + 0.025887349620461464, + 0.009053601883351803, + 0.007866271771490574, + 0.011292815208435059, + 0.004820770584046841, + 0.006196578964591026, + 0.008141634054481983, + 0.005269899033010006, + 0.009936856105923653, + 0.008059276267886162, + 0.007407156750559807, + 0.009145030751824379, + 0.004231806378811598, + 0.006088971626013517, + 0.008939431048929691, + 0.00582386227324605, + 0.007635725662112236, + 0.007055813446640968, + 0.003563752630725503, + 0.008188962005078793, + 0.0027585062198340893, + 0.003989930264651775, + 0.020072678104043007, + 0.003878980875015259, + 0.020484844222664833, + 0.01266842894256115, + 0.008708517998456955, + 0.017130045220255852, + 0.013626099564135075, + 0.015260371379554272, + 0.016905339434742928, + 0.007602047175168991, + 0.017897464334964752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.020578688010573387 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.062027670443058014 + ], + "opposite_logprob_tokens": [ + -2.8120276927948 + ], + "logprob_reward": -0.062027670443058014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062027670443058014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380026936531067, + 0.06222054734826088, + 0.05546707659959793, + 0.12899786233901978, + 0.09360326081514359, + 0.2085649073123932, + 0.08225806802511215, + 0.0632903054356575, + 0.10338504612445831, + 0.06421017646789551 + ], + "credit_map": [ + 0.010297003202140331, + 0.009776326827704906, + 0.01773487590253353, + 0.012292481027543545, + 0.026225216686725616, + 0.009925671853125095, + 0.022440603002905846, + 0.029310518875718117, + 0.007068069651722908, + 0.00776855880394578, + 0.013749120756983757, + 0.0063336496241390705, + 0.0090379873290658, + 0.0036495118401944637, + 0.00937065202742815, + 0.005242996383458376, + 0.007718915119767189, + 0.006655042991042137, + 0.01036669872701168, + 0.009306984953582287, + 0.004264536313712597, + 0.00466192839667201, + 0.008145200088620186, + 0.004347772803157568, + 0.01611783169209957, + 0.01557904202491045, + 0.01765669323503971, + 0.015126424841582775, + 0.01853032223880291, + 0.022061381489038467, + 0.012095417827367783, + 0.01183075550943613, + 0.01295868493616581, + 0.012815873138606548, + 0.015596025623381138, + 0.014709606766700745, + 0.007366800215095282, + 0.011681667529046535, + 0.009115966968238354, + 0.009358628652989864, + 0.017440473660826683, + 0.021529512479901314, + 0.021320391446352005, + 0.030402036383748055, + 0.027061449363827705, + 0.02426171489059925, + 0.03556312993168831, + 0.030986182391643524, + 0.00963077787309885, + 0.011984179727733135, + 0.008422055281698704, + 0.008994145318865776, + 0.01450546644628048, + 0.009150050580501556, + 0.010126912035048008, + 0.009444479830563068, + 0.006615260615944862, + 0.0087904566898942, + 0.008560922928154469, + 0.005487024784088135, + 0.007839597761631012, + 0.008784011006355286, + 0.005982677917927504, + 0.011230356059968472, + 0.01000253763049841, + 0.008641987107694149, + 0.010001759976148605, + 0.006293625570833683, + 0.007848584093153477, + 0.026767805218696594, + 0.00721324048936367, + 0.026615502312779427, + 0.006494004279375076, + 0.006444084923714399, + 0.008547843433916569, + 0.007800258230417967, + 0.010890954174101353, + 0.0055729784071445465, + 0.008259871043264866, + 0.010200181044638157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.062027670443058014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a level angle, slowly and steadily pulls back, gradually revealing a vast bamboo forest shrouded in thick morning mist. At the center of the frame, an elderly man with a youthful face, his long, silvery hair neatly tied into a Taoist bun, clad in a loose-fitting, pure white silk training suit embellished with traditional handmade buttons. He is immersed in a Tai Chi practice, his arms tracing smooth, flowing arcs, his body turning smoothly with the shifting of his center of gravity. Each movement is imbued with inner strength and unhurried rhythm. Soft morning light penetrates the mist and the tall, emerald green bamboo stalks, casting distinct beams of light in the air, casting a subtle shift of light and shadow upon the elderly man and the leaf-strewn ground around him. The scene's dominant color palette is the cool emerald green of the bamboo forest and the milky white of the mist, creating a tranquil, ethereal, and Zen-like atmosphere imbued with Eastern aesthetics.", + "seed": 15156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15156_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": -0.031059269979596137, + "vlm_logprob": -0.031059269979596137, + "vlm_yes_no_margin": 3.824999952316284, + "vlm_token_credit": -0.031059269979596137, + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.02344144694507122 + ], + "opposite_logprob_tokens": [ + -4.023441314697266 + ], + "logprob_reward": -0.02344144694507122, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02344144694507122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023660380393266678, + 0.02554190531373024, + 0.020485930144786835, + 0.0125700943171978, + 0.5939310789108276, + 0.26133209466934204, + 0.02631464973092079, + 0.014246049337089062, + 0.01130647212266922, + 0.010611344128847122 + ], + "credit_map": [ + 0.003406244795769453, + 0.0030353094916790724, + 0.004409195389598608, + 0.0024907265324145555, + 0.0025492608547210693, + 0.0026048338040709496, + 0.002669326728209853, + 0.002495482796803117, + 0.003113508690148592, + 0.0027870049234479666, + 0.004708906169980764, + 0.0026404447853565216, + 0.003332471940666437, + 0.00397952925413847, + 0.002409584354609251, + 0.002570453565567732, + 0.003162662498652935, + 0.0029609212651848793, + 0.0022779700811952353, + 0.0017056716606020927, + 0.002589771756902337, + 0.0023744991049170494, + 0.0021396642550826073, + 0.003274768590927124, + 0.0013557092752307653, + 0.001388991717249155, + 0.002460184507071972, + 0.0011116358218714595, + 0.0014760115882381797, + 0.001921962946653366, + 0.0014702561311423779, + 0.0013853426789864898, + 0.10954100638628006, + 0.045526184141635895, + 0.054226163774728775, + 0.040640126913785934, + 0.05013442784547806, + 0.07440247386693954, + 0.16289347410202026, + 0.0565672442317009, + 0.07110182195901871, + 0.01285728719085455, + 0.012874864973127842, + 0.011186779476702213, + 0.013874510303139687, + 0.02906261757016182, + 0.09272505342960358, + 0.017649151384830475, + 0.0035876077599823475, + 0.0029295168351382017, + 0.003446747548878193, + 0.003574883099645376, + 0.003578376490622759, + 0.002800963819026947, + 0.0047094072215259075, + 0.0016871481202542782, + 0.0017701862379908562, + 0.002212862018495798, + 0.0022300772834569216, + 0.0018230394925922155, + 0.002201819559559226, + 0.0011522335698828101, + 0.0015278274659067392, + 0.001328003709204495, + 0.0013322250451892614, + 0.0015906307380646467, + 0.0016883134376257658, + 0.0008722402271814644, + 0.0017314498545601964, + 0.0013788699870929122, + 0.0013863106723874807, + 0.0013264321023598313, + 0.0011406001867726445, + 0.0007544883992522955, + 0.0017423953395336866, + 0.0010401180479675531, + 0.0011665463680401444, + 0.0022194664925336838, + 0.0012446107575669885, + 0.0013031178386881948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.02344144694507122 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.016099924221634865 + ], + "opposite_logprob_tokens": [ + -4.14109992980957 + ], + "logprob_reward": -0.016099924221634865, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016099924221634865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07715463638305664, + 0.07967735081911087, + 0.18852925300598145, + 0.028982864692807198, + 0.10353383421897888, + 0.32276085019111633, + 0.04535293951630592, + 0.05273312330245972, + 0.06240659952163696, + 0.03886852413415909 + ], + "credit_map": [ + 0.010003522038459778, + 0.010997764766216278, + 0.01080001424998045, + 0.011006156913936138, + 0.010231719352304935, + 0.008681285195052624, + 0.008408077992498875, + 0.007026094011962414, + 0.010809158906340599, + 0.00885740201920271, + 0.012423627078533173, + 0.01077524945139885, + 0.0076311612501740456, + 0.010314339771866798, + 0.009397659450769424, + 0.009468755684792995, + 0.03106391616165638, + 0.02486351691186428, + 0.0214583370834589, + 0.016915515065193176, + 0.02140948548913002, + 0.02059124782681465, + 0.027222251519560814, + 0.025004977360367775, + 0.003264855360612273, + 0.003479551523923874, + 0.0043958439491689205, + 0.0022333385422825813, + 0.002747804392129183, + 0.00400262326002121, + 0.00421834085136652, + 0.004640507511794567, + 0.014198959805071354, + 0.013882488943636417, + 0.0198625810444355, + 0.00925535336136818, + 0.012392954900860786, + 0.01242817658931017, + 0.01064284611493349, + 0.010870477184653282, + 0.08456915616989136, + 0.016014358028769493, + 0.02049945294857025, + 0.014817834831774235, + 0.018520914018154144, + 0.03614075481891632, + 0.10967381298542023, + 0.022524558007717133, + 0.004922534339129925, + 0.007334824185818434, + 0.0060492828488349915, + 0.0047053387388587, + 0.007581974845379591, + 0.005519902799278498, + 0.005391143728047609, + 0.003847941290587187, + 0.011443191207945347, + 0.005534327123314142, + 0.006395149976015091, + 0.004005729220807552, + 0.0033214485738426447, + 0.005355986300855875, + 0.01403570082038641, + 0.0026415896136313677, + 0.007236485835164785, + 0.009079008363187313, + 0.01197288278490305, + 0.004215800669044256, + 0.00518590584397316, + 0.013058213517069817, + 0.005624582525342703, + 0.006033720914274454, + 0.003542956430464983, + 0.004895374178886414, + 0.006340391468256712, + 0.009744682349264622, + 0.003418679116293788, + 0.005783509463071823, + 0.002404689323157072, + 0.0027382408734411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.016099924221634865 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.014200588688254356 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200588688254356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200588688254356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1494714319705963, + 0.08923858404159546, + 0.03715703263878822, + 0.06654878705739975, + 0.17946048080921173, + 0.394528865814209, + 0.024094948545098305, + 0.023640869185328484, + 0.015523988753557205, + 0.020334986969828606 + ], + "credit_map": [ + 0.017228106036782265, + 0.015272368676960468, + 0.020834296941757202, + 0.014502034522593021, + 0.02333688922226429, + 0.023357560858130455, + 0.015974799171090126, + 0.01896537095308304, + 0.013496161438524723, + 0.010949567891657352, + 0.02515297196805477, + 0.007744693197309971, + 0.008113726042211056, + 0.006588438525795937, + 0.008708504028618336, + 0.008484521880745888, + 0.005971960257738829, + 0.0048135267570614815, + 0.0031640473753213882, + 0.0044198292307555676, + 0.004470656160265207, + 0.0038636657409369946, + 0.004775356967002153, + 0.005677991081029177, + 0.0073074051178991795, + 0.007012310437858105, + 0.01566513441503048, + 0.005774081684648991, + 0.006721860263496637, + 0.008800525218248367, + 0.007652250584214926, + 0.007615222595632076, + 0.03017936274409294, + 0.01634848304092884, + 0.02119835466146469, + 0.01265939511358738, + 0.016560731455683708, + 0.02194453403353691, + 0.04226813465356827, + 0.01830149069428444, + 0.105850450694561, + 0.020203914493322372, + 0.020801827311515808, + 0.017606502398848534, + 0.021478798240423203, + 0.04440172761678696, + 0.13769856095314026, + 0.026487044990062714, + 0.0032787169329822063, + 0.0036014921497553587, + 0.0028644728008657694, + 0.003119760425761342, + 0.0022397569846361876, + 0.0034682515542954206, + 0.0033167675137519836, + 0.0022057327441871166, + 0.0024313151370733976, + 0.0034859119914472103, + 0.0037936647422611713, + 0.0027822863776236773, + 0.00392939941957593, + 0.002919825492426753, + 0.0025099420454353094, + 0.0017885224660858512, + 0.0016131517477333546, + 0.0026246095076203346, + 0.0029023918323218822, + 0.0012580185430124402, + 0.0016596103087067604, + 0.002155829221010208, + 0.001768763060681522, + 0.0015416137175634503, + 0.00217061978764832, + 0.0025301966816186905, + 0.0027126541826874018, + 0.0016264478908851743, + 0.002060388680547476, + 0.003429368371143937, + 0.0026110224425792694, + 0.0031942883506417274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014200588688254356 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08680332452058792, + 0.042155634611845016, + 0.03636181727051735, + 0.05285288393497467, + 0.37015461921691895, + 0.06333211064338684, + 0.12326929718255997, + 0.09296204894781113, + 0.04897910729050636, + 0.08312913030385971 + ], + "credit_map": [ + 0.011494109407067299, + 0.00783507339656353, + 0.008926305919885635, + 0.008092173375189304, + 0.010167291387915611, + 0.01887544058263302, + 0.010903692804276943, + 0.010509238578379154, + 0.005620400886982679, + 0.007763660978525877, + 0.004392863716930151, + 0.0032095324713736773, + 0.006304423324763775, + 0.006676021963357925, + 0.0031219597440212965, + 0.0050667766481637955, + 0.0053083947859704494, + 0.005345963407307863, + 0.0039484696462750435, + 0.0027708609122782946, + 0.005873660556972027, + 0.003968304023146629, + 0.0036101231817156076, + 0.005536044016480446, + 0.0059617673978209496, + 0.006142343394458294, + 0.006454722490161657, + 0.0036536259576678276, + 0.006409008521586657, + 0.008797231130301952, + 0.005822557024657726, + 0.009611628949642181, + 0.06740586459636688, + 0.02866087481379509, + 0.03302314504981041, + 0.02602655068039894, + 0.03296670317649841, + 0.04478729888796806, + 0.10021528601646423, + 0.0370689332485199, + 0.013475989922881126, + 0.006859438959509134, + 0.007656068075448275, + 0.0034384680911898613, + 0.005218217149376869, + 0.006753745023161173, + 0.01530707161873579, + 0.004623115994036198, + 0.014608636498451233, + 0.02448270097374916, + 0.018848299980163574, + 0.012864718213677406, + 0.013525031507015228, + 0.01643901690840721, + 0.015483889728784561, + 0.007017003837972879, + 0.013628468848764896, + 0.010958259925246239, + 0.01930672861635685, + 0.013403448276221752, + 0.00721326656639576, + 0.008529962040483952, + 0.015363813377916813, + 0.004558108747005463, + 0.0047859190963208675, + 0.008619029074907303, + 0.009741979651153088, + 0.008559087291359901, + 0.003505292348563671, + 0.007207241374999285, + 0.004074912052601576, + 0.0024856526870280504, + 0.009871360845863819, + 0.01191815547645092, + 0.02072838880121708, + 0.009837279096245766, + 0.006355822551995516, + 0.013032661750912666, + 0.006764643359929323, + 0.004620826803147793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.012537825852632523 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.08901656419038773 + ], + "opposite_logprob_tokens": [ + -2.4640166759490967 + ], + "logprob_reward": -0.08901656419038773, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901656419038773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07167615741491318, + 0.03594671189785004, + 0.02195538394153118, + 0.04642396420240402, + 0.24224896728992462, + 0.42872220277786255, + 0.04958022013306618, + 0.05617231875658035, + 0.03207703307271004, + 0.015197032131254673 + ], + "credit_map": [ + 0.008675849996507168, + 0.009782381355762482, + 0.0102852089330554, + 0.007202583365142345, + 0.009739134460687637, + 0.01200671773403883, + 0.006622090470045805, + 0.007362188771367073, + 0.0038578403182327747, + 0.006506692618131638, + 0.004875109530985355, + 0.0029618642292916775, + 0.005733777768909931, + 0.005061475094407797, + 0.00278659095056355, + 0.004163363482803106, + 0.0036224627401679754, + 0.0035920480731874704, + 0.0021962153259664774, + 0.0017278377199545503, + 0.003467960748821497, + 0.0021094901021569967, + 0.002227633260190487, + 0.0030117349233478308, + 0.005233564414083958, + 0.004126741085201502, + 0.011590413749217987, + 0.003692263038828969, + 0.004600013140588999, + 0.006040317937731743, + 0.0059226336888968945, + 0.005218013655394316, + 0.044772274792194366, + 0.019233684986829758, + 0.022201763466000557, + 0.01655980572104454, + 0.020844705402851105, + 0.030061116442084312, + 0.06518834829330444, + 0.023387271910905838, + 0.11352799087762833, + 0.022655364125967026, + 0.02481055073440075, + 0.019975682720541954, + 0.023480385541915894, + 0.04787875711917877, + 0.14734585583209991, + 0.029047613963484764, + 0.007584639824926853, + 0.005421357695013285, + 0.00593932531774044, + 0.005456656217575073, + 0.006329111754894257, + 0.004675511736422777, + 0.00991251040250063, + 0.00426110764965415, + 0.006858430337160826, + 0.008628170005977154, + 0.008550223894417286, + 0.005766388028860092, + 0.007053032051771879, + 0.008110813796520233, + 0.0056670051999390125, + 0.005538256838917732, + 0.002818251959979534, + 0.006051255390048027, + 0.005326814018189907, + 0.003455640282481909, + 0.003268658649176359, + 0.004473919048905373, + 0.0035571225453168154, + 0.003125371178612113, + 0.002292549703270197, + 0.0015681918011978269, + 0.0022071250714361668, + 0.0009837648831307888, + 0.0014575952664017677, + 0.0031480093020945787, + 0.0013766477350145578, + 0.002163148019462824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.08901656419038773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a level angle, slowly and steadily pulls back, gradually revealing a vast bamboo forest shrouded in thick morning mist. At the center of the frame, an elderly man with a youthful face, his long, silvery hair neatly tied into a Taoist bun, clad in a loose-fitting, pure white silk training suit embellished with traditional handmade buttons. He is immersed in a Tai Chi practice, his arms tracing smooth, flowing arcs, his body turning smoothly with the shifting of his center of gravity. Each movement is imbued with inner strength and unhurried rhythm. Soft morning light penetrates the mist and the tall, emerald green bamboo stalks, casting distinct beams of light in the air, casting a subtle shift of light and shadow upon the elderly man and the leaf-strewn ground around him. The scene's dominant color palette is the cool emerald green of the bamboo forest and the milky white of the mist, creating a tranquil, ethereal, and Zen-like atmosphere imbued with Eastern aesthetics.", + "seed": 15157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_7/index_8287_seed_15157_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": -0.04046864416450262, + "vlm_logprob": -0.04046864416450262, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04046864416450262, + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.0252374317497015 + ], + "opposite_logprob_tokens": [ + -3.9002373218536377 + ], + "logprob_reward": -0.0252374317497015, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0252374317497015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17724812030792236, + 0.08601296693086624, + 0.05460081249475479, + 0.35559067130088806, + 0.030955670401453972, + 0.04622523486614227, + 0.07654629647731781, + 0.043408412486314774, + 0.06126987561583519, + 0.06814193725585938 + ], + "credit_map": [ + 0.022332966327667236, + 0.03421907126903534, + 0.0198751799762249, + 0.015409810468554497, + 0.034413259476423264, + 0.015982048586010933, + 0.020342973992228508, + 0.014672805555164814, + 0.01100713200867176, + 0.007653732784092426, + 0.012918232940137386, + 0.008609263226389885, + 0.011455858126282692, + 0.010376120917499065, + 0.017774200066924095, + 0.00621842872351408, + 0.008278947323560715, + 0.008062181062996387, + 0.004625468049198389, + 0.005535178352147341, + 0.006807444151490927, + 0.0071758185513317585, + 0.004163345787674189, + 0.009952430613338947, + 0.03722984343767166, + 0.04201243072748184, + 0.0347091443836689, + 0.03483101725578308, + 0.028271710500121117, + 0.09175845235586166, + 0.029012074694037437, + 0.057765983045101166, + 0.0037853620015084743, + 0.0044885193929076195, + 0.004452200140804052, + 0.004048943519592285, + 0.004563079681247473, + 0.004153713583946228, + 0.0021604460198432207, + 0.0033034062944352627, + 0.006694257725030184, + 0.00605033291503787, + 0.004765725694596767, + 0.00477921636775136, + 0.006491104606539011, + 0.0056812032125890255, + 0.004948701709508896, + 0.006814689375460148, + 0.0048075891099870205, + 0.003731200937181711, + 0.003447579452767968, + 0.003799147205427289, + 0.007682959083467722, + 0.03096621483564377, + 0.004134229850023985, + 0.017977377399802208, + 0.0034057830926030874, + 0.007024205755442381, + 0.004632601514458656, + 0.004334303084760904, + 0.009296450763940811, + 0.006140692159533501, + 0.004845836665481329, + 0.0037285371217876673, + 0.013615909963846207, + 0.005501268897205591, + 0.012373039498925209, + 0.004995871800929308, + 0.005316014867275953, + 0.005609576124697924, + 0.003827636828646064, + 0.010030556470155716, + 0.007490915711969137, + 0.004964135121554136, + 0.01173978578299284, + 0.003896683920174837, + 0.009746131487190723, + 0.004304602276533842, + 0.017163027077913284, + 0.008836661465466022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0252374317497015 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.012577144429087639 + ], + "opposite_logprob_tokens": [ + -4.387577056884766 + ], + "logprob_reward": -0.012577144429087639, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012577144429087639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08218513429164886, + 0.062167976051568985, + 0.060498397797346115, + 0.35468852519989014, + 0.09196208417415619, + 0.10356992483139038, + 0.0345403216779232, + 0.0381346233189106, + 0.12679345905780792, + 0.04545960947871208 + ], + "credit_map": [ + 0.008323876187205315, + 0.014639648608863354, + 0.010836167261004448, + 0.00682446314021945, + 0.017559189349412918, + 0.008743899874389172, + 0.010234229266643524, + 0.005023658741265535, + 0.008192612789571285, + 0.01043736282736063, + 0.008100823499262333, + 0.0074130333960056305, + 0.008858229033648968, + 0.005433982703834772, + 0.00913508515805006, + 0.004596846643835306, + 0.008362592197954655, + 0.006817012093961239, + 0.005428789649158716, + 0.00545914564281702, + 0.009550637565553188, + 0.011598271317780018, + 0.003239125246182084, + 0.010042817331850529, + 0.031796619296073914, + 0.08130764961242676, + 0.039436761289834976, + 0.051802244037389755, + 0.036928240209817886, + 0.04246241971850395, + 0.037964776158332825, + 0.03298983350396156, + 0.008399010635912418, + 0.013869229704141617, + 0.00771719217300415, + 0.010602022521197796, + 0.0066501977853477, + 0.017663003876805305, + 0.006531778257340193, + 0.020529653877019882, + 0.014539091847836971, + 0.007798361591994762, + 0.00827118381857872, + 0.007260655984282494, + 0.00896074902266264, + 0.03243834897875786, + 0.009019022807478905, + 0.01528251264244318, + 0.0033636153675615788, + 0.0066999224945902824, + 0.004670371301472187, + 0.005179537460207939, + 0.0031202349346131086, + 0.005189329385757446, + 0.002601929008960724, + 0.0037153793964535, + 0.002780296839773655, + 0.0057799662463366985, + 0.004670342896133661, + 0.0036013703793287277, + 0.009221173822879791, + 0.005132467020303011, + 0.0033883503638207912, + 0.0035606538876891136, + 0.012081746011972427, + 0.020474717020988464, + 0.014858382754027843, + 0.020952725782990456, + 0.01589326374232769, + 0.01622544229030609, + 0.01202258002012968, + 0.014284593053162098, + 0.005960628855973482, + 0.00434494623914361, + 0.006269631441682577, + 0.005180560052394867, + 0.006603653077036142, + 0.004309576470404863, + 0.006788978353142738, + 0.006001636385917664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.012577144429087639 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06202609837055206, + 0.02482367306947708, + 0.11738382279872894, + 0.16532133519649506, + 0.10795066505670547, + 0.19236689805984497, + 0.11429696530103683, + 0.056979093700647354, + 0.11693397909402847, + 0.04191751033067703 + ], + "credit_map": [ + 0.00659036124125123, + 0.007121769245713949, + 0.006725619081407785, + 0.005915299989283085, + 0.010754134505987167, + 0.007627641316503286, + 0.010988336987793446, + 0.006302931345999241, + 0.003277029376477003, + 0.003866865299642086, + 0.0023965106811374426, + 0.0023696627467870712, + 0.003774565877392888, + 0.00255765113979578, + 0.004803041461855173, + 0.0017783449729904532, + 0.012392625212669373, + 0.006638533901423216, + 0.0089378971606493, + 0.0046882410533726215, + 0.008863413706421852, + 0.036592159420251846, + 0.003965429496020079, + 0.03530550375580788, + 0.014770139940083027, + 0.022828683257102966, + 0.013591842725872993, + 0.016875838860869408, + 0.01388633158057928, + 0.04373394697904587, + 0.0113492077216506, + 0.028285328298807144, + 0.01844831369817257, + 0.010768634267151356, + 0.008584799244999886, + 0.008515502326190472, + 0.015551628544926643, + 0.016003083437681198, + 0.006729310378432274, + 0.023349370807409286, + 0.0270041786134243, + 0.02093580551445484, + 0.017237279564142227, + 0.0215061753988266, + 0.011884882114827633, + 0.053016554564237595, + 0.01638665795326233, + 0.024395357817411423, + 0.006992575246840715, + 0.007218104787170887, + 0.00462273508310318, + 0.00464092381298542, + 0.01176447607576847, + 0.04543452337384224, + 0.006110721733421087, + 0.027512898668646812, + 0.0060426597483456135, + 0.007559994701296091, + 0.008249393664300442, + 0.006262097042053938, + 0.00916202086955309, + 0.006831600330770016, + 0.007280340883880854, + 0.005590979475528002, + 0.017051581293344498, + 0.01127135381102562, + 0.024011651054024696, + 0.013051138259470463, + 0.011440190486609936, + 0.01115138828754425, + 0.012701751664280891, + 0.016254914924502373, + 0.004888678900897503, + 0.0038411556743085384, + 0.005499327089637518, + 0.0027721934020519257, + 0.007085470482707024, + 0.0032425273675471544, + 0.010225390084087849, + 0.004362758714705706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.014202469028532505 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.02329145185649395 + ], + "opposite_logprob_tokens": [ + -3.7732913494110107 + ], + "logprob_reward": -0.02329145185649395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329145185649395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17236514389514923, + 0.0450923815369606, + 0.11558011919260025, + 0.12127415835857391, + 0.07727578282356262, + 0.14325250685214996, + 0.036478519439697266, + 0.0513346903026104, + 0.0758780911564827, + 0.16146856546401978 + ], + "credit_map": [ + 0.02189651131629944, + 0.038476407527923584, + 0.026753485202789307, + 0.017138881608843803, + 0.01789690926671028, + 0.0209855567663908, + 0.018211515620350838, + 0.011005889624357224, + 0.0053795101121068, + 0.00528579531237483, + 0.00552284624427557, + 0.0043388549238443375, + 0.005824403837323189, + 0.005895229056477547, + 0.009470237419009209, + 0.0033755050972104073, + 0.01151170302182436, + 0.008999316021800041, + 0.007439208682626486, + 0.005798713769763708, + 0.010627402924001217, + 0.03329789265990257, + 0.00530956219881773, + 0.032596319913864136, + 0.011335001327097416, + 0.007276689633727074, + 0.00556987477466464, + 0.0085058668628335, + 0.00983149278908968, + 0.042948391288518906, + 0.006759123411029577, + 0.029047716408967972, + 0.009087090380489826, + 0.00760917691513896, + 0.006627881433814764, + 0.006910212337970734, + 0.007823359221220016, + 0.0173087939620018, + 0.005450705531984568, + 0.016458556056022644, + 0.02566131390631199, + 0.012874562293291092, + 0.013637649826705456, + 0.01300484873354435, + 0.019331462681293488, + 0.021073298528790474, + 0.016020923852920532, + 0.021648429334163666, + 0.0028271363116800785, + 0.0027996140997856855, + 0.001838735188357532, + 0.00212492817081511, + 0.0036028658505529165, + 0.012399274855852127, + 0.002910505747422576, + 0.007975461892783642, + 0.0037930323742330074, + 0.008347729220986366, + 0.004906218033283949, + 0.004998655989766121, + 0.011559662409126759, + 0.007534685544669628, + 0.005390422884374857, + 0.004804285243153572, + 0.012212786823511124, + 0.008088324218988419, + 0.010911502875387669, + 0.0071644531562924385, + 0.008515394292771816, + 0.010303588584065437, + 0.00731347082182765, + 0.01136857084929943, + 0.017942287027835846, + 0.012815715745091438, + 0.028292899951338768, + 0.009815874509513378, + 0.020865632221102715, + 0.012807553634047508, + 0.035896409302949905, + 0.023032190278172493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02329145185649395 + }, + { + "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 long shot featuring an elderly man with long silvery hair tied in a Taoist bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose-fitting pure white silk training suit with traditional handmade buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi with smooth, flowing arm arcs and body turns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast bamboo forest with thick morning mist and emerald green stalks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily pull back to reveal 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": [ + -0.1270347237586975 + ], + "opposite_logprob_tokens": [ + -2.1270346641540527 + ], + "logprob_reward": -0.1270347237586975, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270347237586975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049608390778303146, + 0.034913066774606705, + 0.20242035388946533, + 0.1695118099451065, + 0.13937340676784515, + 0.12415996193885803, + 0.03088083118200302, + 0.051523298025131226, + 0.08030448108911514, + 0.11730436980724335 + ], + "credit_map": [ + 0.0061372630298137665, + 0.007157221436500549, + 0.005918991286307573, + 0.004789971746504307, + 0.007601592224091291, + 0.0053618792444467545, + 0.007391186896711588, + 0.005250285379588604, + 0.003989994991570711, + 0.004833004903048277, + 0.004084773361682892, + 0.004856820683926344, + 0.005454435478895903, + 0.003457791870459914, + 0.004869532771408558, + 0.0033667145762592554, + 0.026106873527169228, + 0.013083168305456638, + 0.013696430251002312, + 0.007359283044934273, + 0.014797426760196686, + 0.0627991333603859, + 0.006217930465936661, + 0.05836009606719017, + 0.011743341572582722, + 0.027478141710162163, + 0.013373618014156818, + 0.018601370975375175, + 0.014454117976129055, + 0.043377362191677094, + 0.011514121666550636, + 0.028969738632440567, + 0.012372753582894802, + 0.016074540093541145, + 0.010126695036888123, + 0.013586685992777348, + 0.014498448930680752, + 0.024808010086417198, + 0.00922708585858345, + 0.03867919370532036, + 0.01865869015455246, + 0.01171091664582491, + 0.006615766324102879, + 0.0099214231595397, + 0.007817168720066547, + 0.040342941880226135, + 0.009224559180438519, + 0.01986849680542946, + 0.0024370686151087284, + 0.0026089148595929146, + 0.0018202709034085274, + 0.0025504459626972675, + 0.0024780903477221727, + 0.01023771706968546, + 0.002386547392234206, + 0.006361775565892458, + 0.004291650839149952, + 0.008471316657960415, + 0.006295436527580023, + 0.004776619840413332, + 0.010715932585299015, + 0.008180229924619198, + 0.0038546547293663025, + 0.004937456455081701, + 0.008686024695634842, + 0.011790916323661804, + 0.010851218365132809, + 0.01191012654453516, + 0.009315834380686283, + 0.008691182360053062, + 0.009016712196171284, + 0.010042470879852772, + 0.009988988749682903, + 0.011481528170406818, + 0.017052195966243744, + 0.007794376462697983, + 0.0208144411444664, + 0.006216796115040779, + 0.03478020057082176, + 0.009175848215818405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.1270347237586975 + } + ], + "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/0000017/group_8/index_tensor([6947, 6947])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_tensor([6947, 6947])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..15effb9833bec3c36f34ea87d2da9f9b6324c04b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_tensor([6947, 6947])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera unfolds from a high angle, beginning with a slow, majestic pullback, bathing the scene in a classic sepia hue reminiscent of vintage photography. Initially, the frame is filled with the mighty Colorado River, winding its way through a vast, arid landscape; the water's surface is a dark, sleek ribbon reflecting soft light. As the camera steadily and deliberately retreats, the water's vast, serpentine curve gradually reveals itself, showcasing how it has carved its ancient path through the dramatic canyon walls. These massive rock formations, with deeply etched strata and weathered textures, are defined by a stark combination of light and long, extended shadows, suggesting a sun low beyond the frame. The slow, continuous camera movement enhances the epic scale, transforming the landscape into a breathtaking topography of geological history and imbuing the silent, expansive vista with a profound and overwhelming atmosphere of timeless grandeur.", + "seed": 26231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_6947_seed_26231_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": -0.06633377955295146, + "vlm_logprob": -0.06633377955295146, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.06633377955295146, + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016274457797408104 + ], + "opposite_logprob_tokens": [ + -4.891274452209473 + ], + "logprob_reward": -0.016274457797408104, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.016274457797408104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12645813822746277, + 0.03423859179019928, + 0.030231285840272903, + 0.11504027247428894, + 0.04720354452729225, + 0.0922459289431572, + 0.10633177310228348, + 0.2026805877685547, + 0.12187539786100388, + 0.1236945167183876 + ], + "credit_map": [ + 0.017114778980612755, + 0.01639804244041443, + 0.02041059546172619, + 0.01263094786554575, + 0.013010768219828606, + 0.022294912487268448, + 0.012074320577085018, + 0.012523761950433254, + 0.0038857338950037956, + 0.009388416074216366, + 0.005301821511238813, + 0.005688402336090803, + 0.0027721470687538385, + 0.0022323783487081528, + 0.0028000890742987394, + 0.0021695990581065416, + 0.0034917579032480717, + 0.005167913157492876, + 0.003830023342743516, + 0.003987196832895279, + 0.00333039159886539, + 0.004325692541897297, + 0.003486673114821315, + 0.0026116350200027227, + 0.009928872808814049, + 0.01215651910752058, + 0.015436168760061264, + 0.015684913843870163, + 0.023815499618649483, + 0.017347920686006546, + 0.010287534445524216, + 0.010382839478552341, + 0.004797779489308596, + 0.009144635871052742, + 0.006257792469114065, + 0.00739440368488431, + 0.006593823432922363, + 0.005083692260086536, + 0.004598132334649563, + 0.0033332814928144217, + 0.008128396235406399, + 0.011340456083416939, + 0.010995970107614994, + 0.014605611562728882, + 0.013624753803014755, + 0.013777229934930801, + 0.009269562549889088, + 0.010503945872187614, + 0.010473594069480896, + 0.017257580533623695, + 0.011485113762319088, + 0.017333203926682472, + 0.012593764811754227, + 0.013232417404651642, + 0.012494748458266258, + 0.011461338959634304, + 0.009322485886514187, + 0.00781951192766428, + 0.011725232936441898, + 0.011956855654716492, + 0.0749322697520256, + 0.06625295430421829, + 0.011798079125583172, + 0.008873187005519867, + 0.005407649092376232, + 0.007888653315603733, + 0.01852444000542164, + 0.007732853293418884, + 0.026462208479642868, + 0.044949475675821304, + 0.0047265393659472466, + 0.006183570716530085, + 0.008713466115295887, + 0.0151415029540658, + 0.012262330390512943, + 0.01983455754816532, + 0.016914697363972664, + 0.014178402721881866, + 0.024175873026251793, + 0.012473681941628456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.016274457797408104 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900078594684601 + ], + "opposite_logprob_tokens": [ + -2.57900071144104 + ], + "logprob_reward": -0.07900078594684601, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900078594684601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07710415124893188, + 0.07233574241399765, + 0.20058362185955048, + 0.205168679356575, + 0.17976464331150055, + 0.026057634502649307, + 0.07177309691905975, + 0.05171005800366402, + 0.04999103397130966, + 0.06551133841276169 + ], + "credit_map": [ + 0.00936021376401186, + 0.010100556537508965, + 0.010613488964736462, + 0.010228719562292099, + 0.0070761023089289665, + 0.011937789618968964, + 0.009813180193305016, + 0.007974096573889256, + 0.00848248228430748, + 0.013993090018630028, + 0.015785593539476395, + 0.008124210871756077, + 0.007721031550318003, + 0.006559268105775118, + 0.005423763766884804, + 0.006246303673833609, + 0.028320571407675743, + 0.01464151218533516, + 0.030115142464637756, + 0.01895870827138424, + 0.02251785434782505, + 0.014414926990866661, + 0.05396736040711403, + 0.017647551372647285, + 0.01849285140633583, + 0.0213159266859293, + 0.029679737985134125, + 0.03113243170082569, + 0.041105709969997406, + 0.028198255226016045, + 0.018617430701851845, + 0.016626333817839622, + 0.01932181417942047, + 0.019462216645479202, + 0.02626589871942997, + 0.030002083629369736, + 0.025881687179207802, + 0.01703795976936817, + 0.018746398389339447, + 0.02304658107459545, + 0.002777526620775461, + 0.0031918371096253395, + 0.0036105746403336525, + 0.0034785927273333073, + 0.003888324135914445, + 0.0035804277285933495, + 0.0030149584636092186, + 0.0025153944734483957, + 0.004380738362669945, + 0.009318135678768158, + 0.009841788560152054, + 0.01873200573027134, + 0.007773526944220066, + 0.007024428807199001, + 0.008441990241408348, + 0.006260480731725693, + 0.0038420376367866993, + 0.0030461414717137814, + 0.005416771862655878, + 0.005882234778255224, + 0.013428888283669949, + 0.011170710436999798, + 0.004528177436441183, + 0.004395098891109228, + 0.003420068183913827, + 0.006795251276344061, + 0.007985597476363182, + 0.009193419478833675, + 0.0066329361870884895, + 0.006969050969928503, + 0.005343073513358831, + 0.0036516361869871616, + 0.0047991713508963585, + 0.005962743889540434, + 0.006135461386293173, + 0.010898256674408913, + 0.012011089362204075, + 0.009403117932379246, + 0.009981340728700161, + 0.0063201552256941795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.07900078594684601 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00220104050822556 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.00220104050822556, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00220104050822556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17857030034065247, + 0.16998209059238434, + 0.07280293852090836, + 0.04245944321155548, + 0.13975508511066437, + 0.045527685433626175, + 0.04623275250196457, + 0.08708338439464569, + 0.1589517891407013, + 0.05863455310463905 + ], + "credit_map": [ + 0.015317863784730434, + 0.021099600940942764, + 0.02637050673365593, + 0.017156725749373436, + 0.02489584870636463, + 0.03131165727972984, + 0.01829533651471138, + 0.02412276715040207, + 0.019021296873688698, + 0.014466403052210808, + 0.04305659234523773, + 0.01755284145474434, + 0.03166688606142998, + 0.01498450618237257, + 0.016341349110007286, + 0.01289222203195095, + 0.014255916699767113, + 0.007617895491421223, + 0.01204304862767458, + 0.009531183168292046, + 0.007168263662606478, + 0.008168971166014671, + 0.008526147343218327, + 0.00549151748418808, + 0.003919907845556736, + 0.005894153378903866, + 0.005137960892170668, + 0.0080290287733078, + 0.005392803810536861, + 0.005300585180521011, + 0.0050163655541837215, + 0.0037686394061893225, + 0.01199896540492773, + 0.015872512012720108, + 0.0186074897646904, + 0.020392809063196182, + 0.02542646788060665, + 0.019965583458542824, + 0.013885837979614735, + 0.013605417683720589, + 0.004322629421949387, + 0.005255653522908688, + 0.005829344969242811, + 0.008632143959403038, + 0.006392465438693762, + 0.004462265409529209, + 0.004901479929685593, + 0.005731701385229826, + 0.0036088679917156696, + 0.006417987402528524, + 0.004338607657700777, + 0.0071067181415855885, + 0.0063256267458200455, + 0.007679980248212814, + 0.005771068390458822, + 0.004983896855264902, + 0.005860072560608387, + 0.004982187878340483, + 0.00664451252669096, + 0.00930663663893938, + 0.025110866874456406, + 0.022741202265024185, + 0.006401874125003815, + 0.006036029662936926, + 0.006551953963935375, + 0.009838612750172615, + 0.02419857867062092, + 0.0095521230250597, + 0.03587241470813751, + 0.0612093061208725, + 0.004418290685862303, + 0.0073105162009596825, + 0.00769060431048274, + 0.009710872545838356, + 0.005563471466302872, + 0.006177968345582485, + 0.006936053745448589, + 0.006306993309408426, + 0.009063961915671825, + 0.007184626068919897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00220104050822556 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07755251973867416, + 0.05926438421010971, + 0.08317962288856506, + 0.04575581103563309, + 0.10226457566022873, + 0.05081918463110924, + 0.02860337868332863, + 0.22231552004814148, + 0.09143908321857452, + 0.23880593478679657 + ], + "credit_map": [ + 0.010460655204951763, + 0.011319343000650406, + 0.010891161859035492, + 0.007015171926468611, + 0.0061636571772396564, + 0.009357575327157974, + 0.012326592579483986, + 0.010018355213105679, + 0.006587967276573181, + 0.010158148594200611, + 0.0094447648152709, + 0.007136374246329069, + 0.006692240480333567, + 0.007037879899144173, + 0.005682648625224829, + 0.006524354685097933, + 0.014954452402889729, + 0.007687451783567667, + 0.012503262609243393, + 0.0064235287718474865, + 0.00688095623627305, + 0.00918115396052599, + 0.015869727358222008, + 0.009679081849753857, + 0.003743884153664112, + 0.006144416984170675, + 0.004899019841104746, + 0.006110339891165495, + 0.006675025448203087, + 0.006651830859482288, + 0.004206935875117779, + 0.007324356585741043, + 0.010137629695236683, + 0.01582987979054451, + 0.01182417944073677, + 0.012598030269145966, + 0.017137087881565094, + 0.013654307462275028, + 0.009669030085206032, + 0.011414432898163795, + 0.004577833693474531, + 0.0069074188359081745, + 0.005842022132128477, + 0.004741067066788673, + 0.010468012653291225, + 0.009009483270347118, + 0.004586992785334587, + 0.004686351865530014, + 0.0023781459312886, + 0.002592730103060603, + 0.0026045532431453466, + 0.0028979144990444183, + 0.004957631230354309, + 0.0056839496828615665, + 0.004591784439980984, + 0.0028966693207621574, + 0.010062012821435928, + 0.009296350181102753, + 0.013009280897676945, + 0.01667359657585621, + 0.08121722936630249, + 0.07006756216287613, + 0.01300696562975645, + 0.008982507511973381, + 0.004866879433393478, + 0.006141000892966986, + 0.014959912747144699, + 0.006438460201025009, + 0.01899798959493637, + 0.031709372997283936, + 0.003316197544336319, + 0.005009261891245842, + 0.03286447003483772, + 0.03243308514356613, + 0.020679550245404243, + 0.0236612968146801, + 0.03745537996292114, + 0.035730425268411636, + 0.02522294595837593, + 0.030758783221244812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.008639458566904068 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555315494537354 + ], + "opposite_logprob_tokens": [ + -1.6005531549453735 + ], + "logprob_reward": -0.22555315494537354, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555315494537354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08666583150625229, + 0.05094857141375542, + 0.06567144393920898, + 0.05623912438750267, + 0.10054010897874832, + 0.09162093698978424, + 0.20655518770217896, + 0.22643215954303741, + 0.061025314033031464, + 0.054301369935274124 + ], + "credit_map": [ + 0.013015766628086567, + 0.010275455191731453, + 0.012001355178654194, + 0.008685044012963772, + 0.009396355599164963, + 0.013831165619194508, + 0.010308765806257725, + 0.009151914156973362, + 0.004987322725355625, + 0.00855353195220232, + 0.006273538805544376, + 0.007051660213619471, + 0.006395820062607527, + 0.006611216813325882, + 0.004884944763034582, + 0.006190530490130186, + 0.009048213250935078, + 0.005376213695853949, + 0.004883827641606331, + 0.005380305927246809, + 0.01572813093662262, + 0.010053383186459541, + 0.009085459634661674, + 0.006115902680903673, + 0.004886307287961245, + 0.006027766037732363, + 0.005786399357020855, + 0.005521408282220364, + 0.014219339936971664, + 0.008375037461519241, + 0.005905880127102137, + 0.005516984965652227, + 0.008209963329136372, + 0.016382629051804543, + 0.012262890115380287, + 0.013980420306324959, + 0.015688331797719002, + 0.012070261873304844, + 0.01192787941545248, + 0.010017722845077515, + 0.011311858892440796, + 0.011937698349356651, + 0.012393499724566936, + 0.011040900833904743, + 0.012118722312152386, + 0.011478545144200325, + 0.008532939478754997, + 0.012806764803826809, + 0.015275422483682632, + 0.019539566710591316, + 0.019032228738069534, + 0.03152835741639137, + 0.039442021399736404, + 0.03480483591556549, + 0.031533896923065186, + 0.015398835763335228, + 0.009418879635632038, + 0.009488013572990894, + 0.013172800652682781, + 0.015696421265602112, + 0.08508921414613724, + 0.07287964969873428, + 0.013291222043335438, + 0.007395932450890541, + 0.002459318144246936, + 0.004125512670725584, + 0.009507915005087852, + 0.005256003234535456, + 0.012797307223081589, + 0.021902723237872124, + 0.002206729957833886, + 0.002769801765680313, + 0.0037784590385854244, + 0.006316750776022673, + 0.005197199061512947, + 0.007700163871049881, + 0.008934255689382553, + 0.006363525986671448, + 0.010522800497710705, + 0.005488207563757896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.22555315494537354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera unfolds from a high angle, beginning with a slow, majestic pullback, bathing the scene in a classic sepia hue reminiscent of vintage photography. Initially, the frame is filled with the mighty Colorado River, winding its way through a vast, arid landscape; the water's surface is a dark, sleek ribbon reflecting soft light. As the camera steadily and deliberately retreats, the water's vast, serpentine curve gradually reveals itself, showcasing how it has carved its ancient path through the dramatic canyon walls. These massive rock formations, with deeply etched strata and weathered textures, are defined by a stark combination of light and long, extended shadows, suggesting a sun low beyond the frame. The slow, continuous camera movement enhances the epic scale, transforming the landscape into a breathtaking topography of geological history and imbuing the silent, expansive vista with a profound and overwhelming atmosphere of timeless grandeur.", + "seed": 26232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_6947_seed_26232_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": -0.07165849308948964, + "vlm_logprob": -0.07165849308948964, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.07165849308948964, + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024913305416703224 + ], + "opposite_logprob_tokens": [ + -4.024913311004639 + ], + "logprob_reward": -0.024913305416703224, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024913305416703224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15075407922267914, + 0.1285470873117447, + 0.0567002147436142, + 0.04123108834028244, + 0.07624877989292145, + 0.028924666345119476, + 0.09088180959224701, + 0.07737304270267487, + 0.23451076447963715, + 0.11482858657836914 + ], + "credit_map": [ + 0.01911150850355625, + 0.019911494106054306, + 0.01945950835943222, + 0.015061015263199806, + 0.046642426401376724, + 0.010952369309961796, + 0.011233382858335972, + 0.008382362313568592, + 0.023348446935415268, + 0.012548976577818394, + 0.03679990768432617, + 0.01191867608577013, + 0.015037987381219864, + 0.007078077178448439, + 0.007271474692970514, + 0.014543531462550163, + 0.011965726502239704, + 0.00448659248650074, + 0.016240470111370087, + 0.0036201386246830225, + 0.0028608404099941254, + 0.004782793112099171, + 0.0048072426579892635, + 0.007936403155326843, + 0.005645881872624159, + 0.006730600725859404, + 0.007600425276905298, + 0.0025902679190039635, + 0.003828849643468857, + 0.004853027872741222, + 0.0035470996517688036, + 0.006434928625822067, + 0.01174132153391838, + 0.006681056227535009, + 0.017112163826823235, + 0.003488802118226886, + 0.003928242716938257, + 0.00907653197646141, + 0.008299686945974827, + 0.015920966863632202, + 0.004912014584988356, + 0.0029971168842166662, + 0.004820529371500015, + 0.00205416651442647, + 0.0040698884986341, + 0.0026919760275632143, + 0.00497543765231967, + 0.0024035342503339052, + 0.01715865544974804, + 0.014943845570087433, + 0.011536499485373497, + 0.006518835201859474, + 0.011554548516869545, + 0.009694859385490417, + 0.011065967381000519, + 0.008408588357269764, + 0.009777727536857128, + 0.008012031204998493, + 0.012045374140143394, + 0.00555385323241353, + 0.008281305432319641, + 0.007309801410883665, + 0.021344400942325592, + 0.005048534367233515, + 0.030321219936013222, + 0.020938675850629807, + 0.04179657995700836, + 0.010378397069871426, + 0.017167024314403534, + 0.023729275912046432, + 0.05928495153784752, + 0.03089461475610733, + 0.013904020190238953, + 0.015736160799860954, + 0.01426523644477129, + 0.008502017706632614, + 0.008191866800189018, + 0.01200195774435997, + 0.028829477727413177, + 0.013397837057709694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.024913305416703224 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203342333436012 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203342333436012, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018203342333436012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0746302530169487, + 0.4302021563053131, + 0.0499875582754612, + 0.009801887907087803, + 0.08135294169187546, + 0.04535109922289848, + 0.05356272682547569, + 0.08357453346252441, + 0.10808257013559341, + 0.06345423310995102 + ], + "credit_map": [ + 0.010409538634121418, + 0.01075038593262434, + 0.01723961904644966, + 0.01001385785639286, + 0.008365686051547527, + 0.005846632178872824, + 0.006490301340818405, + 0.005514233838766813, + 0.08135218918323517, + 0.04520133510231972, + 0.13355104625225067, + 0.03142579272389412, + 0.0322566032409668, + 0.02491568960249424, + 0.02803484909236431, + 0.05346465855836868, + 0.008299103938043118, + 0.007230040617287159, + 0.006856321357190609, + 0.005742741283029318, + 0.007859227247536182, + 0.003691555932164192, + 0.0045960890129208565, + 0.005712474696338177, + 0.0015355271752923727, + 0.0013584938133135438, + 0.0011310824193060398, + 0.0009828433394432068, + 0.0013506260002031922, + 0.0011955046793445945, + 0.0012495458358898759, + 0.0009982644114643335, + 0.01380299124866724, + 0.006896521430462599, + 0.019927235320210457, + 0.0033564423210918903, + 0.0029086617287248373, + 0.008997388184070587, + 0.00797785073518753, + 0.0174858458340168, + 0.005358857102692127, + 0.003767669200897217, + 0.0057631684467196465, + 0.0027528926730155945, + 0.0026851834263652563, + 0.005107495468109846, + 0.014678247272968292, + 0.005237584933638573, + 0.011372278444468975, + 0.008415897376835346, + 0.006646515801548958, + 0.004517309367656708, + 0.005815890617668629, + 0.00660019600763917, + 0.0050030299462378025, + 0.00519160833209753, + 0.009603696875274181, + 0.00683158403262496, + 0.012565952725708485, + 0.003928820136934519, + 0.00767391687259078, + 0.009551909752190113, + 0.026853563264012337, + 0.006565098650753498, + 0.013342440128326416, + 0.008188577368855476, + 0.018440721556544304, + 0.005059471819549799, + 0.007294580340385437, + 0.012031303718686104, + 0.02904278039932251, + 0.014682700857520103, + 0.007147538010030985, + 0.009072466753423214, + 0.005450937431305647, + 0.004865550436079502, + 0.006709040608257055, + 0.007832717150449753, + 0.01577594317495823, + 0.006600042339414358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.018203342333436012 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010451575508341193 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010451575508341193, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010451575508341193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07690468430519104, + 0.08131905645132065, + 0.10203215479850769, + 0.04301833361387253, + 0.21751125156879425, + 0.04858294874429703, + 0.04448229447007179, + 0.19231680035591125, + 0.16449850797653198, + 0.02933395467698574 + ], + "credit_map": [ + 0.011211651377379894, + 0.010081673972308636, + 0.009787649847567081, + 0.011445858515799046, + 0.010144347324967384, + 0.008096279576420784, + 0.008301572874188423, + 0.007835648953914642, + 0.01060524769127369, + 0.010380661115050316, + 0.015350142493844032, + 0.013120321556925774, + 0.011018854565918446, + 0.005947405938059092, + 0.006591358222067356, + 0.008305054157972336, + 0.02319474332034588, + 0.007085076998919249, + 0.03056526742875576, + 0.0062126959674060345, + 0.005680887494236231, + 0.007971511222422123, + 0.007507858797907829, + 0.013814103789627552, + 0.006067164707928896, + 0.007690349128097296, + 0.00672954972833395, + 0.0032371734268963337, + 0.004153888672590256, + 0.0050309933722019196, + 0.004103911574929953, + 0.006005299277603626, + 0.035579390823841095, + 0.01869973912835121, + 0.05321831628680229, + 0.009507795795798302, + 0.009305644780397415, + 0.02427702397108078, + 0.021170703694224358, + 0.045752622187137604, + 0.0058126808144152164, + 0.005292830057442188, + 0.004917987622320652, + 0.0035863725934177637, + 0.004488486330956221, + 0.005588545463979244, + 0.013373956084251404, + 0.005522087216377258, + 0.005961344111710787, + 0.007297029253095388, + 0.005026258062571287, + 0.0039146458730101585, + 0.006340545602142811, + 0.005651767831295729, + 0.005034362431615591, + 0.005256337579339743, + 0.02350858598947525, + 0.010635285638272762, + 0.03491741791367531, + 0.0071943108923733234, + 0.009380613453686237, + 0.023081831634044647, + 0.06803864240646362, + 0.015560115687549114, + 0.020966699346899986, + 0.013429922983050346, + 0.02778599224984646, + 0.0072741140611469746, + 0.009167938493192196, + 0.018517933785915375, + 0.045191485434770584, + 0.02216440439224243, + 0.0037834446411579847, + 0.004438308533281088, + 0.0035363684874027967, + 0.002203904092311859, + 0.0024714674800634384, + 0.003360169241204858, + 0.007022086065262556, + 0.0025182052049785852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0010451575508341193 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205299124121666 + ], + "opposite_logprob_tokens": [ + -2.8120529651641846 + ], + "logprob_reward": -0.06205299124121666, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205299124121666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10067243129014969, + 0.08156982064247131, + 0.04921070858836174, + 0.04084335267543793, + 0.1669381856918335, + 0.04024053364992142, + 0.06512614339590073, + 0.09201665967702866, + 0.3376568555831909, + 0.02572532743215561 + ], + "credit_map": [ + 0.011718645691871643, + 0.011482120491564274, + 0.020208708941936493, + 0.015136506408452988, + 0.013625799678266048, + 0.010642784647643566, + 0.00929520558565855, + 0.008562647737562656, + 0.010698636062443256, + 0.01344320923089981, + 0.012566193006932735, + 0.010221143253147602, + 0.010454721748828888, + 0.007321543525904417, + 0.008114399388432503, + 0.008749962784349918, + 0.007671532221138477, + 0.006766792852431536, + 0.009273044764995575, + 0.003877747105434537, + 0.004951711278408766, + 0.004917055368423462, + 0.005562277976423502, + 0.0061905402690172195, + 0.005820722319185734, + 0.005500469356775284, + 0.009009324014186859, + 0.0018798509845510125, + 0.003183607244864106, + 0.004263360518962145, + 0.003857682691887021, + 0.007328331936150789, + 0.02565864659845829, + 0.018579699099063873, + 0.037735678255558014, + 0.0077371192164719105, + 0.010724488645792007, + 0.01645292155444622, + 0.016451535746455193, + 0.033598076552152634, + 0.005373252090066671, + 0.0029814669396728277, + 0.007464177906513214, + 0.0018986407667398453, + 0.0028446712531149387, + 0.004141609184443951, + 0.011583941988646984, + 0.003952767699956894, + 0.01189590897411108, + 0.009998752735555172, + 0.007967607118189335, + 0.005415662657469511, + 0.004780549090355635, + 0.006048699375241995, + 0.010813509114086628, + 0.008205447345972061, + 0.011578346602618694, + 0.0052805026061832905, + 0.016196617856621742, + 0.003259160090237856, + 0.004926307592540979, + 0.010775672271847725, + 0.03154003992676735, + 0.008460000157356262, + 0.04323985427618027, + 0.024690307676792145, + 0.059399280697107315, + 0.014643242582678795, + 0.01833522692322731, + 0.037625011056661606, + 0.0928332656621933, + 0.046890631318092346, + 0.002513193292543292, + 0.003922831732779741, + 0.003378965426236391, + 0.00205999705940485, + 0.00228706793859601, + 0.0025668973103165627, + 0.006065992172807455, + 0.0029303801711648703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.06205299124121666 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520776689052582 + ], + "opposite_logprob_tokens": [ + -1.5020776987075806 + ], + "logprob_reward": -0.2520776689052582, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520776689052582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1378694772720337, + 0.06266294419765472, + 0.12186910212039948, + 0.07196599990129471, + 0.2531612813472748, + 0.08657853305339813, + 0.0470997653901577, + 0.14004918932914734, + 0.06494833528995514, + 0.013795409351587296 + ], + "credit_map": [ + 0.016229860484600067, + 0.016345810145139694, + 0.01892928220331669, + 0.013914227485656738, + 0.038465943187475204, + 0.012226475402712822, + 0.012635225430130959, + 0.00912265945225954, + 0.010754185728728771, + 0.006193835288286209, + 0.018329467624425888, + 0.006401543039828539, + 0.004577186889946461, + 0.0038202418945729733, + 0.004145931452512741, + 0.008440554141998291, + 0.018511245027184486, + 0.017793618142604828, + 0.020694250240921974, + 0.016688251867890358, + 0.012618915177881718, + 0.011237372644245625, + 0.014152825810015202, + 0.010172617621719837, + 0.011313200928270817, + 0.007750064134597778, + 0.0162360779941082, + 0.003948863595724106, + 0.004297229461371899, + 0.007784726098179817, + 0.006694961339235306, + 0.013940873555839062, + 0.041175585240125656, + 0.02144509181380272, + 0.061576712876558304, + 0.010484903119504452, + 0.009182739071547985, + 0.030158426612615585, + 0.02480939030647278, + 0.054328422993421555, + 0.015360253863036633, + 0.011607765220105648, + 0.009989692829549313, + 0.006832062266767025, + 0.007993288338184357, + 0.010719327256083488, + 0.014147206209599972, + 0.00992893148213625, + 0.006916282698512077, + 0.007237564772367477, + 0.005195565987378359, + 0.0038959530647844076, + 0.008645625784993172, + 0.0055567254312336445, + 0.004780184011906385, + 0.004871869459748268, + 0.016846012324094772, + 0.008144674822688103, + 0.02455926313996315, + 0.005751049146056175, + 0.0067477780394256115, + 0.017404425889253616, + 0.048060446977615356, + 0.012535547837615013, + 0.007897160947322845, + 0.005653261672705412, + 0.011024652980268002, + 0.003548222128301859, + 0.005987546872347593, + 0.0070511577650904655, + 0.015467222779989243, + 0.008319107815623283, + 0.0018610891420394182, + 0.003328880062326789, + 0.0016336095286533237, + 0.0015229511773213744, + 0.0014212867245078087, + 0.0012747959699481726, + 0.0017275601858273149, + 0.0010252358624711633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.2520776689052582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera unfolds from a high angle, beginning with a slow, majestic pullback, bathing the scene in a classic sepia hue reminiscent of vintage photography. Initially, the frame is filled with the mighty Colorado River, winding its way through a vast, arid landscape; the water's surface is a dark, sleek ribbon reflecting soft light. As the camera steadily and deliberately retreats, the water's vast, serpentine curve gradually reveals itself, showcasing how it has carved its ancient path through the dramatic canyon walls. These massive rock formations, with deeply etched strata and weathered textures, are defined by a stark combination of light and long, extended shadows, suggesting a sun low beyond the frame. The slow, continuous camera movement enhances the epic scale, transforming the landscape into a breathtaking topography of geological history and imbuing the silent, expansive vista with a profound and overwhelming atmosphere of timeless grandeur.", + "seed": 26233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_6947_seed_26233_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": -0.05189092170912772, + "vlm_logprob": -0.05189092170912772, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.05189092170912772, + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009559450671076775 + ], + "opposite_logprob_tokens": [ + -5.634559631347656 + ], + "logprob_reward": -0.009559450671076775, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.009559450671076775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25330352783203125, + 0.09152574092149734, + 0.1330244541168213, + 0.07079150527715683, + 0.028154006227850914, + 0.02968377061188221, + 0.06814435124397278, + 0.09531932324171066, + 0.14319804310798645, + 0.08685527741909027 + ], + "credit_map": [ + 0.035335149616003036, + 0.024996759369969368, + 0.036077361553907394, + 0.028373785316944122, + 0.03201639652252197, + 0.04295562207698822, + 0.019815076142549515, + 0.033733390271663666, + 0.014601515606045723, + 0.00888555683195591, + 0.014874438755214214, + 0.011387250386178493, + 0.009342363104224205, + 0.010906514711678028, + 0.013112992979586124, + 0.00841511320322752, + 0.023555897176265717, + 0.013389213010668755, + 0.01940864324569702, + 0.012204768136143684, + 0.019481277093291283, + 0.020606402307748795, + 0.010373779572546482, + 0.014004470780491829, + 0.01192664448171854, + 0.005930421873927116, + 0.0117948604747653, + 0.006680172402411699, + 0.008071045391261578, + 0.011554356664419174, + 0.007997729815542698, + 0.006836273241788149, + 0.004387225024402142, + 0.0031382099259644747, + 0.0047179097309708595, + 0.005072711035609245, + 0.002809347817674279, + 0.0028492894489318132, + 0.00307249347679317, + 0.0021068204659968615, + 0.005014466121792793, + 0.0033354556653648615, + 0.004545397125184536, + 0.004754403606057167, + 0.0029698361176997423, + 0.0038884563837200403, + 0.0025347070768475533, + 0.0026410496793687344, + 0.010013117454946041, + 0.006287495139986277, + 0.008326142095029354, + 0.008161364123225212, + 0.007752172648906708, + 0.009861933998763561, + 0.010567951016128063, + 0.007174181751906872, + 0.012131663039326668, + 0.009268887341022491, + 0.00957405287772417, + 0.011407792568206787, + 0.010797951370477676, + 0.021183082833886147, + 0.009684909135103226, + 0.011270981281995773, + 0.01750597357749939, + 0.017470795661211014, + 0.017385952174663544, + 0.016294479370117188, + 0.011467250064015388, + 0.023627320304512978, + 0.013818829320371151, + 0.025627439841628075, + 0.009062755852937698, + 0.00918690674006939, + 0.014061243273317814, + 0.011154256761074066, + 0.006137014366686344, + 0.019431181252002716, + 0.00442021107301116, + 0.013401713222265244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.009559450671076775 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635188400745392 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635188400745392, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635188400745392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15013214945793152, + 0.1619900017976761, + 0.07498672604560852, + 0.06253120303153992, + 0.04377713426947594, + 0.03016289882361889, + 0.03223703056573868, + 0.09239915758371353, + 0.3013196587562561, + 0.050464075058698654 + ], + "credit_map": [ + 0.018397711217403412, + 0.01436989102512598, + 0.016450826078653336, + 0.012886809185147285, + 0.022896086797118187, + 0.03408856689929962, + 0.012417605146765709, + 0.018624644726514816, + 0.019168827682733536, + 0.0256490558385849, + 0.019263526424765587, + 0.026994163170456886, + 0.015183603391051292, + 0.0198574960231781, + 0.019485779106616974, + 0.01638755016028881, + 0.010372824035584927, + 0.010336457751691341, + 0.010904747061431408, + 0.011788874864578247, + 0.008187355473637581, + 0.008675756864249706, + 0.007343275938183069, + 0.0073774345219135284, + 0.010628788731992245, + 0.008594302460551262, + 0.00903585646301508, + 0.007842480204999447, + 0.005777142941951752, + 0.00865070428699255, + 0.006038314662873745, + 0.005963610485196114, + 0.005645631812512875, + 0.005635356996208429, + 0.005924305412918329, + 0.00777400890365243, + 0.0043327766470611095, + 0.004096867050975561, + 0.006692534778267145, + 0.003675656160339713, + 0.0036365697160363197, + 0.003938610665500164, + 0.003980295266956091, + 0.005183367989957333, + 0.0046895635314285755, + 0.0033079618588089943, + 0.0027840316761285067, + 0.0026425004471093416, + 0.004406373482197523, + 0.004210125654935837, + 0.004625842906534672, + 0.004307127092033625, + 0.004877686500549316, + 0.00390962278470397, + 0.0028298394754529, + 0.003070410806685686, + 0.011231010779738426, + 0.013925991952419281, + 0.013284609653055668, + 0.01607581414282322, + 0.008923598565161228, + 0.010351396165788174, + 0.011189672164618969, + 0.0074170660227537155, + 0.02899199165403843, + 0.021648989990353584, + 0.030256226658821106, + 0.03252300247550011, + 0.023419367149472237, + 0.07173126935958862, + 0.02507469803094864, + 0.06767411530017853, + 0.0037260567769408226, + 0.005571205634623766, + 0.005518447607755661, + 0.004709282889962196, + 0.0028255919460207224, + 0.012874734587967396, + 0.003946748562157154, + 0.011292008683085442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02635188400745392 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011812621960416436 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011812621960416436, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011812621960416436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2944549024105072, + 0.05633186176419258, + 0.07919151335954666, + 0.04845923185348511, + 0.04700164869427681, + 0.06327826529741287, + 0.1199059784412384, + 0.0700083076953888, + 0.13446834683418274, + 0.08689997345209122 + ], + "credit_map": [ + 0.040647704154253006, + 0.03623936325311661, + 0.045015376061201096, + 0.03122684545814991, + 0.032471466809511185, + 0.04540157690644264, + 0.028171515092253685, + 0.035281017422676086, + 0.0069696069695055485, + 0.009162629023194313, + 0.0066027455031871796, + 0.006275315769016743, + 0.007886847481131554, + 0.007236028090119362, + 0.005590218119323254, + 0.006608469877392054, + 0.012348187156021595, + 0.008066615089774132, + 0.014275411143898964, + 0.00840801652520895, + 0.010265829972922802, + 0.009966709651052952, + 0.00886362511664629, + 0.006997114047408104, + 0.008680682629346848, + 0.003989135846495628, + 0.006549330428242683, + 0.004532897844910622, + 0.006696249358355999, + 0.007960218004882336, + 0.005251364782452583, + 0.00479934923350811, + 0.008759713731706142, + 0.004699653945863247, + 0.006153610534965992, + 0.006071175914257765, + 0.005564335733652115, + 0.005902142729610205, + 0.005622455384582281, + 0.004228560719639063, + 0.010764529928565025, + 0.006497771944850683, + 0.0068465592339634895, + 0.008321463130414486, + 0.009383979253470898, + 0.007385652977973223, + 0.008602919988334179, + 0.005475388374179602, + 0.018040448427200317, + 0.011757969856262207, + 0.016308149322867393, + 0.013759899884462357, + 0.014950454235076904, + 0.018088340759277344, + 0.015085086226463318, + 0.011915617622435093, + 0.007552914321422577, + 0.005861257202923298, + 0.006518376525491476, + 0.008329971693456173, + 0.007443789392709732, + 0.019706925377249718, + 0.00525464816018939, + 0.009340429678559303, + 0.013901636004447937, + 0.014218349941074848, + 0.02024821937084198, + 0.023150162771344185, + 0.012468310073018074, + 0.01617910899221897, + 0.012187786400318146, + 0.022114761173725128, + 0.009533656761050224, + 0.010151906870305538, + 0.012930135242640972, + 0.011944163590669632, + 0.006205282639712095, + 0.017061030492186546, + 0.006212510634213686, + 0.012861278839409351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0011812621960416436 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203360855579376 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.06203360855579376, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203360855579376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3667737543582916, + 0.08659761399030685, + 0.0438542515039444, + 0.04263870790600777, + 0.07918476313352585, + 0.05032480135560036, + 0.026981672272086143, + 0.05996464192867279, + 0.08912096172571182, + 0.15455876290798187 + ], + "credit_map": [ + 0.05103246867656708, + 0.03131512552499771, + 0.05066753923892975, + 0.029819032177329063, + 0.06066632643342018, + 0.07073772698640823, + 0.030782729387283325, + 0.04175281524658203, + 0.014839793555438519, + 0.008815787732601166, + 0.013085192069411278, + 0.008928420022130013, + 0.011062916368246078, + 0.014087886549532413, + 0.007438892498612404, + 0.008338730782270432, + 0.00690738158300519, + 0.006093665026128292, + 0.004236604552716017, + 0.003889914136379957, + 0.0063912104815244675, + 0.00684939231723547, + 0.005023412872105837, + 0.004462667275220156, + 0.006189247593283653, + 0.005509802605956793, + 0.005371966864913702, + 0.005314436741173267, + 0.005448732990771532, + 0.006499845068901777, + 0.003988189157098532, + 0.0043164887465536594, + 0.013337218202650547, + 0.01272182259708643, + 0.00945452880114317, + 0.01040589064359665, + 0.009167341515421867, + 0.008206134662032127, + 0.009076614864170551, + 0.006815214175730944, + 0.008464301005005836, + 0.007401210721582174, + 0.0064815557561814785, + 0.006096118129789829, + 0.006335568148642778, + 0.005473933182656765, + 0.0053712171502411366, + 0.004700897261500359, + 0.00439076405018568, + 0.0033421069383621216, + 0.0033675620798021555, + 0.0025460589677095413, + 0.0025838979054242373, + 0.005337799899280071, + 0.0025402416940778494, + 0.0028732395730912685, + 0.00832532811909914, + 0.005757227074354887, + 0.006082945037633181, + 0.007611038163304329, + 0.010552770458161831, + 0.008370455354452133, + 0.007475229445844889, + 0.005789644084870815, + 0.010033678263425827, + 0.009794209152460098, + 0.013073856011033058, + 0.012357768602669239, + 0.008587231859564781, + 0.014149040915071964, + 0.011794920079410076, + 0.0093302670866251, + 0.012758221477270126, + 0.013408763334155083, + 0.023705923929810524, + 0.01483269315212965, + 0.010613267309963703, + 0.042327698320150375, + 0.006024245638400316, + 0.03088795393705368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06203360855579376 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032840311527252 + ], + "opposite_logprob_tokens": [ + -1.9103283882141113 + ], + "logprob_reward": -0.16032840311527252, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032840311527252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17271502315998077, + 0.17053917050361633, + 0.197910875082016, + 0.07433879375457764, + 0.057561300694942474, + 0.03358206897974014, + 0.035447973757982254, + 0.07595178484916687, + 0.07647409290075302, + 0.1054789274930954 + ], + "credit_map": [ + 0.021515071392059326, + 0.01556923147290945, + 0.020478829741477966, + 0.014527316205203533, + 0.028489504009485245, + 0.0309592355042696, + 0.013581064529716969, + 0.027594760060310364, + 0.018680062144994736, + 0.026992391794919968, + 0.019536660984158516, + 0.024256641045212746, + 0.03165445476770401, + 0.021504968404769897, + 0.01504984125494957, + 0.012864147312939167, + 0.0325738750398159, + 0.026917919516563416, + 0.029744187369942665, + 0.026982799172401428, + 0.018263371661305428, + 0.025910746306180954, + 0.017810285091400146, + 0.01970769092440605, + 0.012767416425049305, + 0.009146583266556263, + 0.009440421126782894, + 0.01039927825331688, + 0.008905676193535328, + 0.009933670982718468, + 0.006728869862854481, + 0.007016879040747881, + 0.00937710702419281, + 0.008285291492938995, + 0.0060835592448711395, + 0.00824351329356432, + 0.007557658478617668, + 0.006104669068008661, + 0.0068085952661931515, + 0.005100906360894442, + 0.004803049378097057, + 0.003559483215212822, + 0.004015322308987379, + 0.005067880265414715, + 0.004309637937694788, + 0.004813049454241991, + 0.004008792340755463, + 0.0030048529151827097, + 0.00633326405659318, + 0.0037074789870530367, + 0.004594539292156696, + 0.003488676156848669, + 0.0036736612673848867, + 0.005769745446741581, + 0.004416221287101507, + 0.003464385634288192, + 0.010328544303774834, + 0.008907973766326904, + 0.008621849119663239, + 0.011928259395062923, + 0.008361702784895897, + 0.012184986844658852, + 0.008358767256140709, + 0.007259695790708065, + 0.00746126938611269, + 0.009742298163473606, + 0.011035717092454433, + 0.010121408849954605, + 0.009812205098569393, + 0.012019581161439419, + 0.00717106182128191, + 0.00911055225878954, + 0.009375439025461674, + 0.012577286921441555, + 0.015097150579094887, + 0.010772396810352802, + 0.008056276477873325, + 0.024431098252534866, + 0.006044279783964157, + 0.019124995917081833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.16032840311527252 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera unfolds from a high angle, beginning with a slow, majestic pullback, bathing the scene in a classic sepia hue reminiscent of vintage photography. Initially, the frame is filled with the mighty Colorado River, winding its way through a vast, arid landscape; the water's surface is a dark, sleek ribbon reflecting soft light. As the camera steadily and deliberately retreats, the water's vast, serpentine curve gradually reveals itself, showcasing how it has carved its ancient path through the dramatic canyon walls. These massive rock formations, with deeply etched strata and weathered textures, are defined by a stark combination of light and long, extended shadows, suggesting a sun low beyond the frame. The slow, continuous camera movement enhances the epic scale, transforming the landscape into a breathtaking topography of geological history and imbuing the silent, expansive vista with a profound and overwhelming atmosphere of timeless grandeur.", + "seed": 26234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_8/index_6947_seed_26234_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": -0.07640087414183654, + "vlm_logprob": -0.07640087414183654, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.07640087414183654, + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005808498244732618 + ], + "opposite_logprob_tokens": [ + -6.130808353424072 + ], + "logprob_reward": -0.005808498244732618, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005808498244732618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22520571947097778, + 0.045843735337257385, + 0.06182394549250603, + 0.048053715378046036, + 0.06625229865312576, + 0.03999313339591026, + 0.3767412304878235, + 0.036857012659311295, + 0.029541287571191788, + 0.06968798488378525 + ], + "credit_map": [ + 0.023567145690321922, + 0.025983991101384163, + 0.029917234554886818, + 0.05005645006895065, + 0.030395591631531715, + 0.018198804929852486, + 0.028844812884926796, + 0.018241707235574722, + 0.0061553530395030975, + 0.005743000190705061, + 0.004618079401552677, + 0.005767002701759338, + 0.004871348384767771, + 0.00565893342718482, + 0.007716804277151823, + 0.005313210189342499, + 0.013246873393654823, + 0.009300502017140388, + 0.004275388550013304, + 0.004937531892210245, + 0.010486681945621967, + 0.005522505845874548, + 0.008496053516864777, + 0.005558408796787262, + 0.005639195907860994, + 0.005445918068289757, + 0.004975851159542799, + 0.0037425202317535877, + 0.011750767938792706, + 0.004194540902972221, + 0.0071403165347874165, + 0.005164607428014278, + 0.009682531468570232, + 0.008763090707361698, + 0.005728958640247583, + 0.007398937828838825, + 0.007501661311835051, + 0.0057173422537744045, + 0.011645115911960602, + 0.009814659133553505, + 0.00523634348064661, + 0.004557250067591667, + 0.007137098349630833, + 0.0038499447982758284, + 0.005668900907039642, + 0.0037019834853708744, + 0.005398810841143131, + 0.00444280169904232, + 0.10719171911478043, + 0.05104292184114456, + 0.01962457224726677, + 0.015415419824421406, + 0.03812339901924133, + 0.021659554913640022, + 0.035920727998018265, + 0.08776287734508514, + 0.004655409138649702, + 0.0042177001014351845, + 0.005772708915174007, + 0.004881768487393856, + 0.005364250857383013, + 0.002752502914518118, + 0.0049630627036094666, + 0.004249609541147947, + 0.0033883468713611364, + 0.0036418212112039328, + 0.004307287745177746, + 0.002571239834651351, + 0.00310687767341733, + 0.002321020932868123, + 0.005999945569783449, + 0.004204748198390007, + 0.013939372263848782, + 0.012117520906031132, + 0.0070386831648647785, + 0.005333222448825836, + 0.009506145492196083, + 0.005753964651376009, + 0.008786468766629696, + 0.007212605327367783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.005808498244732618 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067402152344584465 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.0067402152344584465, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067402152344584465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2315695434808731, + 0.06823061406612396, + 0.06730130314826965, + 0.04362849146127701, + 0.06410517543554306, + 0.06978902965784073, + 0.21477843821048737, + 0.029405394569039345, + 0.04660409316420555, + 0.16458791494369507 + ], + "credit_map": [ + 0.023649955168366432, + 0.030080832540988922, + 0.025412775576114655, + 0.04277421534061432, + 0.02463233284652233, + 0.021732136607170105, + 0.03780880197882652, + 0.02547849901020527, + 0.010416320525109768, + 0.008973373100161552, + 0.0068881930783391, + 0.007750172168016434, + 0.008635994046926498, + 0.005496032070368528, + 0.011222206987440586, + 0.008848323486745358, + 0.013062209822237492, + 0.01046556793153286, + 0.005717018153518438, + 0.006049905903637409, + 0.008658196777105331, + 0.006038718391209841, + 0.010118558071553707, + 0.00719112902879715, + 0.009232573211193085, + 0.005521839018911123, + 0.0036485728342086077, + 0.0030429763719439507, + 0.007170996628701687, + 0.0034015681594610214, + 0.006784300319850445, + 0.004825666546821594, + 0.006850200705230236, + 0.005212079267948866, + 0.004626637790352106, + 0.0061839669942855835, + 0.01203418243676424, + 0.008054682053625584, + 0.01287698745727539, + 0.008266440592706203, + 0.010505640879273415, + 0.007657515816390514, + 0.010081948712468147, + 0.007363817188888788, + 0.00959747564047575, + 0.004919670056551695, + 0.01090039499104023, + 0.008762567304074764, + 0.054218072444200516, + 0.026917608454823494, + 0.01316041313111782, + 0.009526189416646957, + 0.033110108226537704, + 0.012552584521472454, + 0.02052892930805683, + 0.044764529913663864, + 0.003450226504355669, + 0.00237515801563859, + 0.0035947104915976524, + 0.0027601714245975018, + 0.003953351639211178, + 0.004370932467281818, + 0.005521135870367289, + 0.0033797076903283596, + 0.0046522291377186775, + 0.00418423255905509, + 0.005641743540763855, + 0.0055500417947769165, + 0.0065595051273703575, + 0.005886083003133535, + 0.007235626690089703, + 0.006894631776958704, + 0.03503601253032684, + 0.028400471433997154, + 0.0182962603867054, + 0.014220626093447208, + 0.017227591946721077, + 0.011200498789548874, + 0.023553021252155304, + 0.016653437167406082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0067402152344584465 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00026854246971197426 + ], + "opposite_logprob_tokens": [ + -8.250268936157227 + ], + "logprob_reward": -0.00026854246971197426, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.00026854246971197426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10665971040725708, + 0.04103033244609833, + 0.05276615545153618, + 0.03142174705862999, + 0.04160471260547638, + 0.061327844858169556, + 0.5296632051467896, + 0.018527574837207794, + 0.04159704968333244, + 0.07540174573659897 + ], + "credit_map": [ + 0.010683209635317326, + 0.014627685770392418, + 0.01697465404868126, + 0.01596313901245594, + 0.015757840126752853, + 0.012005151249468327, + 0.01184022519737482, + 0.008807813748717308, + 0.004447667859494686, + 0.007085506804287434, + 0.002982919104397297, + 0.0053604948334395885, + 0.004613985773175955, + 0.004521953873336315, + 0.007121172267943621, + 0.004896633792668581, + 0.008643759414553642, + 0.007070173975080252, + 0.0041486420668661594, + 0.005849550012499094, + 0.008507543243467808, + 0.004854283295571804, + 0.007511848118156195, + 0.006180356722325087, + 0.006329024210572243, + 0.0043130479753017426, + 0.0026495461352169514, + 0.002730631036683917, + 0.005513560958206654, + 0.0033125481568276882, + 0.00369600229896605, + 0.0028773858211934566, + 0.005595394875854254, + 0.005470464006066322, + 0.0037733856588602066, + 0.004983002785593271, + 0.0051811388693749905, + 0.004189938306808472, + 0.00781975220888853, + 0.004591634031385183, + 0.010709021240472794, + 0.008182802237570286, + 0.006423281971365213, + 0.005406459327787161, + 0.008646553382277489, + 0.0037929704412817955, + 0.010608378797769547, + 0.007558375131338835, + 0.14850586652755737, + 0.07092791050672531, + 0.02571268007159233, + 0.02268599160015583, + 0.05567154288291931, + 0.029639579355716705, + 0.052128102630376816, + 0.12439147382974625, + 0.003238692879676819, + 0.002542615169659257, + 0.002598344814032316, + 0.002101082354784012, + 0.002441081451252103, + 0.0014929876197129488, + 0.0022498462349176407, + 0.0018629244295880198, + 0.006665684748440981, + 0.005169144365936518, + 0.007023331709206104, + 0.0053035118617117405, + 0.0043544466607272625, + 0.0039794244803488255, + 0.004735271446406841, + 0.004366237670183182, + 0.013860596343874931, + 0.013115409761667252, + 0.007843922823667526, + 0.005463353358209133, + 0.01115507259964943, + 0.005462876055389643, + 0.008687959983944893, + 0.009812552481889725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.00026854246971197426 + }, + { + "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 dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2593616843223572, + 0.08095689862966537, + 0.050608910620212555, + 0.03368610516190529, + 0.0369190089404583, + 0.1009325459599495, + 0.23615863919258118, + 0.08816491812467575, + 0.04650284722447395, + 0.06670846045017242 + ], + "credit_map": [ + 0.021140387281775475, + 0.038995884358882904, + 0.023439038544893265, + 0.07586244493722916, + 0.02301069349050522, + 0.020951827988028526, + 0.036044564098119736, + 0.01991686224937439, + 0.010940006002783775, + 0.011422039940953255, + 0.006392676383256912, + 0.008089371025562286, + 0.009561443701386452, + 0.010135065764188766, + 0.010995701886713505, + 0.013420597650110722, + 0.010058716870844364, + 0.006577776744961739, + 0.003530168905854225, + 0.003294015536084771, + 0.008048967458307743, + 0.004727572668343782, + 0.007917222566902637, + 0.006454467307776213, + 0.005647579673677683, + 0.003937460016459227, + 0.002738876501098275, + 0.003211002564057708, + 0.003756128717213869, + 0.0029883896932005882, + 0.00639680540189147, + 0.005009863525629044, + 0.005498761776834726, + 0.005191073752939701, + 0.004083994776010513, + 0.0030014400836080313, + 0.004927840083837509, + 0.002615200588479638, + 0.006359786260873079, + 0.005240910686552525, + 0.016611766070127487, + 0.01480687316507101, + 0.008404982276260853, + 0.00900865625590086, + 0.013747351244091988, + 0.00825609639286995, + 0.010240860283374786, + 0.019855966791510582, + 0.046277571469545364, + 0.025664497166872025, + 0.014264230616390705, + 0.011781470850110054, + 0.06176754832267761, + 0.013924713246524334, + 0.02576901763677597, + 0.036709580570459366, + 0.022448595613241196, + 0.012200002558529377, + 0.007973791100084782, + 0.00870093610137701, + 0.0073686097748577595, + 0.005442807916551828, + 0.008406081236898899, + 0.015624088235199451, + 0.007361719850450754, + 0.005657538305968046, + 0.0061409869231283665, + 0.003478822996839881, + 0.005897653754800558, + 0.004081957507878542, + 0.005792624782770872, + 0.00809154100716114, + 0.008233553729951382, + 0.009496801532804966, + 0.008864683099091053, + 0.0063005974516272545, + 0.010420446284115314, + 0.006878488697111607, + 0.008697422221302986, + 0.007816468365490437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "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 scene dominated by the Colorado River winding through an arid landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the water surface depicted as a dark, sleek ribbon reflecting soft light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive rock formations with deeply etched strata and weathered textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the entire scene bathed in a classic sepia hue reminiscent of vintage photography?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a high angle and perform a slow pullback?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860527515411377 + ], + "opposite_logprob_tokens": [ + -1.2236052751541138 + ], + "logprob_reward": -0.34860527515411377, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34860527515411377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11317657679319382, + 0.0669768899679184, + 0.08178297430276871, + 0.033419761806726456, + 0.050891805440187454, + 0.09922806173563004, + 0.41602975130081177, + 0.08139900118112564, + 0.024384627118706703, + 0.032710567116737366 + ], + "credit_map": [ + 0.009832949377596378, + 0.015825672075152397, + 0.012103009968996048, + 0.027513494715094566, + 0.014321810565888882, + 0.010179491713643074, + 0.015526430681347847, + 0.00787370279431343, + 0.010540327988564968, + 0.010015927255153656, + 0.006666670553386211, + 0.008134582079946995, + 0.009651788510382175, + 0.005621979013085365, + 0.00911695696413517, + 0.007228657603263855, + 0.01643700711429119, + 0.011744855903089046, + 0.0060854461044073105, + 0.007938653230667114, + 0.013088439591228962, + 0.0064112343825399876, + 0.011788275092840195, + 0.008289068937301636, + 0.005428098142147064, + 0.004492889624089003, + 0.003890798194333911, + 0.0038105084095150232, + 0.003917634952813387, + 0.002975247800350189, + 0.005099381320178509, + 0.0038052028976380825, + 0.008013376034796238, + 0.006455256137996912, + 0.006279502063989639, + 0.005720462184399366, + 0.0061823781579732895, + 0.004511409904807806, + 0.0064587644301354885, + 0.007270655129104853, + 0.01833626814186573, + 0.012962480075657368, + 0.012523168697953224, + 0.008523129858076572, + 0.013565447181463242, + 0.007452164310961962, + 0.014549019746482372, + 0.011316382326185703, + 0.0663011372089386, + 0.04416821897029877, + 0.024006355553865433, + 0.027330849319696426, + 0.11764004081487656, + 0.02683897688984871, + 0.05442849174141884, + 0.055315665900707245, + 0.019240543246269226, + 0.011479126289486885, + 0.0067887515760958195, + 0.007554508280009031, + 0.007903863675892353, + 0.005035243462771177, + 0.00990548636764288, + 0.013491482473909855, + 0.003812249982729554, + 0.0034514281433075666, + 0.002367164473980665, + 0.0021582446061074734, + 0.003574972739443183, + 0.0024695061147212982, + 0.0037193738389760256, + 0.0028316860552877188, + 0.006148718763142824, + 0.005598966497927904, + 0.0034436893183737993, + 0.00274596712552011, + 0.004176449496299028, + 0.0024851129855960608, + 0.004503785166889429, + 0.003607878927141428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.34860527515411377 + } + ], + "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/0000017/group_9/index_tensor([1438, 1438])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_tensor([1438, 1438])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a1f22e14d6baa70638f07f18bf27b3a33341023d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_tensor([1438, 1438])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, horizontal medium shot focuses on a scene composed entirely of pixels, embodying a retro 8-bit gaming aesthetic. A wooden bookshelf, composed of dark and light brown pixels, dominates the frame, neatly arranged with vibrantly colored books featuring distinct pixelated edges. Their spines display saturated red, emerald green, royal blue, and bright yellow. A simple, light-gray pixelated wall frames the background, casting an even, bright light across the frame, creating virtually no shadows. The action begins with a red book at the end of the queue cautiously tilting to one side before quickly losing balance and crashing into a nearby green book, triggering a chain reaction. Immediately, the books fall one by one in a slightly stuttering, frame-by-frame animation, tumbling and plummeting through the air before disappearing off the bottom of the frame. The entire sequence is imbued with a sense of dynamism and nostalgia. The vibrant colors and unique pixelated dynamics create a lighthearted, playful atmosphere.", + "seed": 28604, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_1438_seed_28604_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": -0.06440275993663817, + "vlm_logprob": -0.06440275993663817, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06440275993663817, + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.1265198290348053 + ], + "opposite_logprob_tokens": [ + -2.6265199184417725 + ], + "logprob_reward": -0.1265198290348053, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1265198290348053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013749206438660622, + 0.008053118363022804, + 0.007504759822040796, + 0.0047584534622728825, + 0.011776277795433998, + 0.01481742411851883, + 0.005042119883000851, + 0.019227975979447365, + 0.014481522142887115, + 0.9005891680717468 + ], + "credit_map": [ + 0.0022096685133874416, + 0.0022790348157286644, + 0.0008034509955905378, + 0.0009398571564815938, + 0.0026919282972812653, + 0.002823710674419999, + 0.0012051792582497, + 0.0007963773095980287, + 0.0009192414581775665, + 0.0010474162409082055, + 0.001161211053840816, + 0.00040263377013616264, + 0.0008579600835219026, + 0.001968987053260207, + 0.0010448701214045286, + 0.0006507994839921594, + 0.0009059046278707683, + 0.0007244726293720305, + 0.0004733734531328082, + 0.00036522699519991875, + 0.001570671098306775, + 0.002014080761000514, + 0.000798596884123981, + 0.0006524333148263395, + 0.0005511902854777873, + 0.00030586685170419514, + 0.00028787372866645455, + 0.0002615417179185897, + 0.0006129750981926918, + 0.0014208657667040825, + 0.0007778549916110933, + 0.00054028513841331, + 0.0030565429478883743, + 0.0038976711221039295, + 0.0006701354286633432, + 0.0008840251248329878, + 0.001616948051378131, + 0.0007290488574653864, + 0.00046066686627455056, + 0.00046123904758132994, + 0.002444186247885227, + 0.0007938275230117142, + 0.001513902796432376, + 0.004366724751889706, + 0.0005491225747391582, + 0.0007751559023745358, + 0.0008776318863965571, + 0.003496871329843998, + 0.0007775809499435127, + 0.00029158388497307897, + 0.0012243884848430753, + 0.0007186639704741538, + 0.0003266840649303049, + 0.00017992117500398308, + 0.0007962597301229835, + 0.0007270380738191307, + 0.0007458874024450779, + 0.0004471165011636913, + 0.006056667305529118, + 0.004727205261588097, + 0.0006703798426315188, + 0.00040461873868480325, + 0.004105732310563326, + 0.002070368966087699, + 0.0006244751857593656, + 0.00043032318353652954, + 0.0047204080037772655, + 0.004144796170294285, + 0.00043290582834742963, + 0.0003240026126150042, + 0.0017366003012284636, + 0.0020680103916674852, + 0.011378192342817783, + 0.006613261066377163, + 0.008684840053319931, + 0.009430330246686935, + 0.01550382375717163, + 0.009624162688851357, + 0.4169082045555115, + 0.4224463403224945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.1265198290348053 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.062093548476696014 + ], + "opposite_logprob_tokens": [ + -2.812093496322632 + ], + "logprob_reward": -0.062093548476696014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062093548476696014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021799305453896523, + 0.008041296154260635, + 0.007657822221517563, + 0.019330676645040512, + 0.039531905204057693, + 0.004275797866284847, + 0.008667631074786186, + 0.033761199563741684, + 0.027053549885749817, + 0.8298807740211487 + ], + "credit_map": [ + 0.003705153474584222, + 0.0043422263115644455, + 0.00176858005579561, + 0.0008575458195991814, + 0.0046946206130087376, + 0.004031725227832794, + 0.0012288617435842752, + 0.0011705935467034578, + 0.0011311968555673957, + 0.0007132720202207565, + 0.001755525590851903, + 0.000515843799803406, + 0.0005911876214668155, + 0.0016503097722306848, + 0.0007084358367137611, + 0.000975524599198252, + 0.001536408206447959, + 0.0008579451241530478, + 0.0012374366633594036, + 0.0005202308529987931, + 0.001227547531016171, + 0.0007920362986624241, + 0.0007331318338401616, + 0.000753086234908551, + 0.0024426712188869715, + 0.0016102947993203998, + 0.001342074479907751, + 0.0012270153965801, + 0.003651887644082308, + 0.004518887493759394, + 0.0024844191502779722, + 0.0020534275099635124, + 0.005934411659836769, + 0.009952545166015625, + 0.003130842698737979, + 0.01101647038012743, + 0.0034485645592212677, + 0.0018542965408414602, + 0.002315671881660819, + 0.001879107323475182, + 0.0005237528821453452, + 0.0004405042855069041, + 0.0003694425686262548, + 0.000525998417288065, + 0.0008631599484942853, + 0.000433241919381544, + 0.0007197812665253878, + 0.0003999168984591961, + 0.0010613197227939963, + 0.0002657913137227297, + 0.0020126812160015106, + 0.0011233132099732757, + 0.0002893626515287906, + 0.00034917175071313977, + 0.0014789880951866508, + 0.0020870042499154806, + 0.0014490255853161216, + 0.000844297930598259, + 0.008375968784093857, + 0.008022953756153584, + 0.0018369328463450074, + 0.0011755606392398477, + 0.006741295102983713, + 0.005315171089023352, + 0.0012078031431883574, + 0.0011671772226691246, + 0.007015347480773926, + 0.008303741924464703, + 0.0009071032400242984, + 0.0004683588049374521, + 0.0047581735998392105, + 0.0032258452847599983, + 0.011191015131771564, + 0.004203629679977894, + 0.007951507344841957, + 0.008698275312781334, + 0.013467303477227688, + 0.008152110502123833, + 0.385109007358551, + 0.39110803604125977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.062093548476696014 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.0013585394481197 + ], + "opposite_logprob_tokens": [ + -6.626358509063721 + ], + "logprob_reward": -0.0013585394481197, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013585394481197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021791908890008926, + 0.006206031888723373, + 0.008873452432453632, + 0.006993759423494339, + 0.02436777763068676, + 0.03398153558373451, + 0.006098499987274408, + 0.013032602146267891, + 0.006218268536031246, + 0.8724361658096313 + ], + "credit_map": [ + 0.0029495644848793745, + 0.0034657444339245558, + 0.0009035792318172753, + 0.0009950930252671242, + 0.007018391974270344, + 0.0031239655800163746, + 0.0020127438474446535, + 0.001322825555689633, + 0.0007363777840510011, + 0.0006997099262662232, + 0.0005505263688974082, + 0.00037264186539687216, + 0.0006492076208814979, + 0.0015154816210269928, + 0.0008300715708173811, + 0.0008520149858668447, + 0.0009379912517033517, + 0.0006231582374311984, + 0.0009158551692962646, + 0.0004665947053581476, + 0.0016762894811108708, + 0.0018773966003209352, + 0.0013265033485367894, + 0.0010496631730347872, + 0.0006641616346314549, + 0.0005891815526410937, + 0.000380182929802686, + 0.0005189948715269566, + 0.0012237349292263389, + 0.0016514388844370842, + 0.0010712294606491923, + 0.0008948355098254979, + 0.006730865221470594, + 0.007471841294318438, + 0.0016685329610481858, + 0.002807182027027011, + 0.0027921253349632025, + 0.0011277437442913651, + 0.0008371858857572079, + 0.0009323023841716349, + 0.0012713954783976078, + 0.0014976462116464972, + 0.001736829406581819, + 0.012775563634932041, + 0.0008782703662291169, + 0.0018281674711033702, + 0.0016912475693970919, + 0.012302414514124393, + 0.0011033741757273674, + 0.0004788333608303219, + 0.0009862546576187015, + 0.0007224146393127739, + 0.00034397406852804124, + 0.0002985422615893185, + 0.0008459216332994401, + 0.001319185015745461, + 0.0009914052207022905, + 0.000547362316865474, + 0.0030308361165225506, + 0.0022059425245970488, + 0.0008796784095466137, + 0.0006050998927094042, + 0.002249837387353182, + 0.0025224401615560055, + 0.00035316767753101885, + 0.0002327060792595148, + 0.002026165835559368, + 0.0013513998128473759, + 0.00031120487255975604, + 0.00016984157264232635, + 0.0008297355379909277, + 0.0009440473513677716, + 0.010908903554081917, + 0.002342505380511284, + 0.008208150044083595, + 0.008342444896697998, + 0.009554549120366573, + 0.008687681518495083, + 0.40927526354789734, + 0.41511669754981995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.0013585394481197 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.06203562393784523 + ], + "opposite_logprob_tokens": [ + -2.81203556060791 + ], + "logprob_reward": -0.06203562393784523, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203562393784523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05429257079958916, + 0.018236516043543816, + 0.01236311998218298, + 0.022241048514842987, + 0.07500714808702469, + 0.04148688167333603, + 0.03051682375371456, + 0.05261992663145065, + 0.05404946580529213, + 0.6391865015029907 + ], + "credit_map": [ + 0.00956827960908413, + 0.010950672440230846, + 0.0017614808166399598, + 0.0016402301844209433, + 0.014116240665316582, + 0.010690230876207352, + 0.0026412932202219963, + 0.0029241370502859354, + 0.002030793111771345, + 0.00237318710424006, + 0.0027363370172679424, + 0.0011496179504320025, + 0.0022564646787941456, + 0.0028733178041875362, + 0.0025389050133526325, + 0.002277892315760255, + 0.0015266640111804008, + 0.0009735471103340387, + 0.0016851889668032527, + 0.0013150488957762718, + 0.001371339545585215, + 0.0018629698315635324, + 0.0018628668040037155, + 0.0017654945841059089, + 0.0020937013905495405, + 0.0016957063926383853, + 0.0010555696208029985, + 0.0012834054650738835, + 0.00402968842536211, + 0.005937732290476561, + 0.0029385306406766176, + 0.003206714056432247, + 0.02043350785970688, + 0.025582656264305115, + 0.004808296449482441, + 0.006053950637578964, + 0.010186411440372467, + 0.0035087703727185726, + 0.0022840185556560755, + 0.002149540465325117, + 0.0024628096725791693, + 0.0019763021264225245, + 0.0026875059120357037, + 0.015321106649935246, + 0.0012151835253462195, + 0.002097246004268527, + 0.00238387961871922, + 0.013342845253646374, + 0.004619005136191845, + 0.0021109541412442923, + 0.004560512490570545, + 0.003434077836573124, + 0.001564342761412263, + 0.0012426057364791632, + 0.0055076912976801395, + 0.007477635517716408, + 0.0024214857257902622, + 0.001619475893676281, + 0.011041687801480293, + 0.01297914981842041, + 0.0032148745376616716, + 0.0018463368760421872, + 0.011499063111841679, + 0.007997850887477398, + 0.0019111502915620804, + 0.0016124428948387504, + 0.0196171086281538, + 0.013839046470820904, + 0.0027959849685430527, + 0.0017401449149474502, + 0.0059822131879627705, + 0.0065513732843101025, + 0.008547850884497166, + 0.0034835513215512037, + 0.009120680391788483, + 0.010793662630021572, + 0.007780976593494415, + 0.007090848870575428, + 0.29351750016212463, + 0.2988514006137848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.06203562393784523 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.07000625878572464 + ], + "opposite_logprob_tokens": [ + -2.6950063705444336 + ], + "logprob_reward": -0.07000625878572464, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000625878572464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036317504942417145, + 0.018408305943012238, + 0.006609932519495487, + 0.013303235173225403, + 0.06831859797239304, + 0.04332569241523743, + 0.01393047533929348, + 0.0703650638461113, + 0.03419356793165207, + 0.6952276229858398 + ], + "credit_map": [ + 0.006576086860150099, + 0.006557331420481205, + 0.0038352059200406075, + 0.0015262519009411335, + 0.006872653029859066, + 0.007220110855996609, + 0.0020161594729870558, + 0.0017137102549895644, + 0.005949388723820448, + 0.0031217799987643957, + 0.0021011114586144686, + 0.0007808065856806934, + 0.0013778534485027194, + 0.003027749713510275, + 0.0011932614725083113, + 0.0008563565206713974, + 0.0008552763611078262, + 0.0010598755907267332, + 0.0009219364728778601, + 0.000503413553815335, + 0.0009248584392480552, + 0.0007016067975200713, + 0.000906478613615036, + 0.0007364872144535184, + 0.0010534875327721238, + 0.000974436174146831, + 0.0008254231652244925, + 0.0006518989684991539, + 0.0021153877023607492, + 0.0035394493024796247, + 0.0018957413267344236, + 0.002247412223368883, + 0.020404621958732605, + 0.020482301712036133, + 0.004599583335220814, + 0.006792244967073202, + 0.007537726312875748, + 0.003513624193146825, + 0.0023578277323395014, + 0.002630671951919794, + 0.002632490824908018, + 0.002119190525263548, + 0.002673534443601966, + 0.01566288247704506, + 0.0016664944123476744, + 0.0022712964564561844, + 0.0029715881682932377, + 0.013328219763934612, + 0.0021338826045393944, + 0.00042180303717032075, + 0.002408660016953945, + 0.0018274388276040554, + 0.00046192025183700025, + 0.0005114845116622746, + 0.002431846922263503, + 0.0037334407679736614, + 0.0025560513604432344, + 0.0019223925191909075, + 0.013506058603525162, + 0.016072191298007965, + 0.004904757719486952, + 0.0034961903002113104, + 0.013182198628783226, + 0.014725231565535069, + 0.001066291704773903, + 0.0008098710677586496, + 0.012844712473452091, + 0.009951836429536343, + 0.000980223761871457, + 0.0006109364912845194, + 0.00405636802315712, + 0.0038733268156647682, + 0.00974686536937952, + 0.0029195649549365044, + 0.006617461796849966, + 0.008147264830768108, + 0.010508454404771328, + 0.0068679079413414, + 0.3229113817214966, + 0.3275087773799896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.07000625878572464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal medium shot focuses on a scene composed entirely of pixels, embodying a retro 8-bit gaming aesthetic. A wooden bookshelf, composed of dark and light brown pixels, dominates the frame, neatly arranged with vibrantly colored books featuring distinct pixelated edges. Their spines display saturated red, emerald green, royal blue, and bright yellow. A simple, light-gray pixelated wall frames the background, casting an even, bright light across the frame, creating virtually no shadows. The action begins with a red book at the end of the queue cautiously tilting to one side before quickly losing balance and crashing into a nearby green book, triggering a chain reaction. Immediately, the books fall one by one in a slightly stuttering, frame-by-frame animation, tumbling and plummeting through the air before disappearing off the bottom of the frame. The entire sequence is imbued with a sense of dynamism and nostalgia. The vibrant colors and unique pixelated dynamics create a lighthearted, playful atmosphere.", + "seed": 28605, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_1438_seed_28605_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": -0.04606386530213058, + "vlm_logprob": -0.04606386530213058, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04606386530213058, + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.09715169668197632 + ], + "opposite_logprob_tokens": [ + -2.972151756286621 + ], + "logprob_reward": -0.09715169668197632, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09715169668197632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30064767599105835, + 0.0705408975481987, + 0.02711230143904686, + 0.018524594604969025, + 0.018789300695061684, + 0.06077877804636955, + 0.028781987726688385, + 0.028588011860847473, + 0.1530931144952774, + 0.29314327239990234 + ], + "credit_map": [ + 0.029833853244781494, + 0.04678519442677498, + 0.042364493012428284, + 0.045204900205135345, + 0.02415473200380802, + 0.04535729065537453, + 0.03119085170328617, + 0.035756368190050125, + 0.007801867090165615, + 0.010501072742044926, + 0.010403383523225784, + 0.009956986643373966, + 0.004976282827556133, + 0.009468101896345615, + 0.007218477316200733, + 0.010214731097221375, + 0.0029630158096551895, + 0.003817093325778842, + 0.003909093793481588, + 0.003175602061673999, + 0.0027463468722999096, + 0.00333491712808609, + 0.0026407677214592695, + 0.004525465425103903, + 0.002650439739227295, + 0.00307046459056437, + 0.0023437594063580036, + 0.0024888571351766586, + 0.0021110987290740013, + 0.0024445997551083565, + 0.0016068060649558902, + 0.0018085695337504148, + 0.002515269909054041, + 0.001701732282526791, + 0.0016291053034365177, + 0.0014065745053812861, + 0.00484324200078845, + 0.0032197581604123116, + 0.0011559997219592333, + 0.0023176176473498344, + 0.0053291176445782185, + 0.004073719028383493, + 0.015903793275356293, + 0.005924669094383717, + 0.015348190441727638, + 0.006447985768318176, + 0.004107962362468243, + 0.0036433436907827854, + 0.0036537779960781336, + 0.004342836327850819, + 0.0033858211245387793, + 0.0028678940143436193, + 0.004405276384204626, + 0.005232764407992363, + 0.0018166835652664304, + 0.003076934488490224, + 0.004004620481282473, + 0.005158018786460161, + 0.002754966262727976, + 0.003337291767820716, + 0.005071826279163361, + 0.003706286195665598, + 0.002031948883086443, + 0.0025230529718101025, + 0.011934145353734493, + 0.04801725223660469, + 0.010389021597802639, + 0.008366098627448082, + 0.03829260170459747, + 0.019796933978796005, + 0.008772754110395908, + 0.007524322252720594, + 0.007181307766586542, + 0.12823833525180817, + 0.00522868474945426, + 0.006785229779779911, + 0.11823641508817673, + 0.015817053616046906, + 0.005722439847886562, + 0.00593382166698575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.09715169668197632 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.008662149310112 + ], + "opposite_logprob_tokens": [ + -4.758662223815918 + ], + "logprob_reward": -0.008662149310112, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662149310112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11731576919555664, + 0.03341200202703476, + 0.04171733185648918, + 0.04598486050963402, + 0.0547623410820961, + 0.04482708498835564, + 0.06334851682186127, + 0.09050630033016205, + 0.14766283333301544, + 0.3604629635810852 + ], + "credit_map": [ + 0.014208974316716194, + 0.01848306693136692, + 0.01952996291220188, + 0.017247123643755913, + 0.007849600166082382, + 0.012379607185721397, + 0.009918899275362492, + 0.01769852824509144, + 0.003982761409133673, + 0.0036645825020968914, + 0.0034830195363610983, + 0.002205737866461277, + 0.0057790279388427734, + 0.004183011129498482, + 0.0042474837973713875, + 0.0058663757517933846, + 0.004178016446530819, + 0.004265897441655397, + 0.004161653574556112, + 0.0044002048671245575, + 0.002338665071874857, + 0.00870567187666893, + 0.0028952707070857286, + 0.010771949775516987, + 0.006012484431266785, + 0.006638689432293177, + 0.005323802120983601, + 0.003504235530272126, + 0.002564233262091875, + 0.010838553309440613, + 0.003143700072541833, + 0.007959160953760147, + 0.007209818810224533, + 0.0068883649073541164, + 0.004670875612646341, + 0.00429275119677186, + 0.007101892028003931, + 0.012085840106010437, + 0.003782575251534581, + 0.008730221539735794, + 0.004361126571893692, + 0.0029511586762964725, + 0.01091405563056469, + 0.003888384671881795, + 0.011049010790884495, + 0.0053115664049983025, + 0.002795472042635083, + 0.0035563106648623943, + 0.009873107075691223, + 0.00441147293895483, + 0.004628999624401331, + 0.005109874531626701, + 0.008905069902539253, + 0.011675475165247917, + 0.004307837691158056, + 0.014436675235629082, + 0.009966514073312283, + 0.01098273042589426, + 0.009090413339436054, + 0.007249615620821714, + 0.018761679530143738, + 0.018018009141087532, + 0.010792353190481663, + 0.0056449854746460915, + 0.015078573487699032, + 0.0323338657617569, + 0.0107218436896801, + 0.011147288605570793, + 0.03162258863449097, + 0.023567331954836845, + 0.010316782630980015, + 0.012874558568000793, + 0.01278511993587017, + 0.14728066325187683, + 0.010232473723590374, + 0.011123710311949253, + 0.1405116766691208, + 0.016230499371886253, + 0.01167129073292017, + 0.01062751840800047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008662149310112 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.0025059746112674475 + ], + "opposite_logprob_tokens": [ + -6.002505779266357 + ], + "logprob_reward": -0.0025059746112674475, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025059746112674475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1237688735127449, + 0.02617744356393814, + 0.03881102427840233, + 0.035462722182273865, + 0.0424354150891304, + 0.030415073037147522, + 0.022902965545654297, + 0.04532906040549278, + 0.13069160282611847, + 0.5040056705474854 + ], + "credit_map": [ + 0.013626177795231342, + 0.0190841406583786, + 0.014284205622971058, + 0.012520006857812405, + 0.01558173168450594, + 0.014257221482694149, + 0.014024943113327026, + 0.02039046585559845, + 0.0030314824543893337, + 0.0034166069235652685, + 0.004545971285551786, + 0.004183975048363209, + 0.0019307613838464022, + 0.0034839066211134195, + 0.002539148787036538, + 0.0030455971136689186, + 0.002883746288716793, + 0.0033377083018422127, + 0.003259937511757016, + 0.0027665914967656136, + 0.007485263515263796, + 0.006508879829198122, + 0.0060333674773573875, + 0.0065355380065739155, + 0.0044539798982441425, + 0.0055005610920488834, + 0.00407437002286315, + 0.0027740418445318937, + 0.0024268205743283033, + 0.004637354053556919, + 0.0032340793404728174, + 0.008361523039638996, + 0.0024103079922497272, + 0.00615990674123168, + 0.0026111931074410677, + 0.002258256310597062, + 0.00944020226597786, + 0.009146771393716335, + 0.001801602658815682, + 0.008607182651758194, + 0.003171712625771761, + 0.0021504932083189487, + 0.005849815439432859, + 0.0021726384293287992, + 0.00926769059151411, + 0.004516206216067076, + 0.0016617199871689081, + 0.001624802127480507, + 0.0023162655998021364, + 0.0028864487539976835, + 0.002055651042610407, + 0.0024489688221365213, + 0.0037167370319366455, + 0.0045811147429049015, + 0.001735436380840838, + 0.0031623474787920713, + 0.006053948309272528, + 0.007251711096614599, + 0.003597348928451538, + 0.004351295065134764, + 0.008286362513899803, + 0.008589551784098148, + 0.003712802194058895, + 0.003486047266051173, + 0.010486806742846966, + 0.04079967364668846, + 0.00812759529799223, + 0.006220737472176552, + 0.03339433670043945, + 0.019287055358290672, + 0.005589407403022051, + 0.006786014419049025, + 0.010279584676027298, + 0.22717581689357758, + 0.005664506461471319, + 0.008679156191647053, + 0.22069339454174042, + 0.017949098721146584, + 0.007251233793795109, + 0.006312998943030834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0025059746112674475 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.07896168529987335 + ], + "opposite_logprob_tokens": [ + -2.5789616107940674 + ], + "logprob_reward": -0.07896168529987335, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896168529987335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10005071014165878, + 0.023575907573103905, + 0.018887339159846306, + 0.017079494893550873, + 0.024034088477492332, + 0.2621731460094452, + 0.034997809678316116, + 0.11248935759067535, + 0.12184108793735504, + 0.28487110137939453 + ], + "credit_map": [ + 0.008879303932189941, + 0.010338366031646729, + 0.010684001259505749, + 0.012828044593334198, + 0.010597198270261288, + 0.018651319667696953, + 0.008144690655171871, + 0.019927777349948883, + 0.0031073675490915775, + 0.00492457440122962, + 0.0027405163273215294, + 0.0019690366461873055, + 0.0018546743085607886, + 0.0031498735770583153, + 0.0018285714322701097, + 0.004001293797045946, + 0.001682766480371356, + 0.0019558786880224943, + 0.0017232770333066583, + 0.0011870646849274635, + 0.0029106461443006992, + 0.0036273906007409096, + 0.002141606295481324, + 0.0036587067879736423, + 0.002480250084772706, + 0.002103836042806506, + 0.0017126306192949414, + 0.0012153512798249722, + 0.0012914937688037753, + 0.0037604081444442272, + 0.0013284579617902637, + 0.0031870664097368717, + 0.0032870429567992687, + 0.0034847643692046404, + 0.0030306389089673758, + 0.002941789571195841, + 0.004324212204664946, + 0.0025936432648450136, + 0.002166393445804715, + 0.002205601194873452, + 0.03832778334617615, + 0.029139867052435875, + 0.03490842878818512, + 0.018106017261743546, + 0.033244337886571884, + 0.05519266054034233, + 0.014194194227457047, + 0.03905985504388809, + 0.004095002543181181, + 0.003121150890365243, + 0.002912037307396531, + 0.0038594265934079885, + 0.005109686870127916, + 0.0061640688218176365, + 0.0021067727357149124, + 0.007629659026861191, + 0.011738674715161324, + 0.015113106928765774, + 0.008164476603269577, + 0.007446440868079662, + 0.028992190957069397, + 0.025630412623286247, + 0.007871880196034908, + 0.007532154675573111, + 0.01116525661200285, + 0.03456761687994003, + 0.010762385092675686, + 0.008541209623217583, + 0.035824134945869446, + 0.009728377684950829, + 0.005320874974131584, + 0.00593122374266386, + 0.006677526980638504, + 0.12147887796163559, + 0.00700029544532299, + 0.005976883228868246, + 0.12276606261730194, + 0.009424063377082348, + 0.006305539980530739, + 0.005241820588707924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.07896168529987335 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.04303782060742378 + ], + "opposite_logprob_tokens": [ + -3.1680378913879395 + ], + "logprob_reward": -0.04303782060742378, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303782060742378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07442492246627808, + 0.04229247197508812, + 0.01109540555626154, + 0.013767442665994167, + 0.028452716767787933, + 0.27336186170578003, + 0.027262933552265167, + 0.059557780623435974, + 0.2122712880373001, + 0.2575131356716156 + ], + "credit_map": [ + 0.009327114559710026, + 0.009740225970745087, + 0.00732449721544981, + 0.006120915059000254, + 0.014878712594509125, + 0.007130710408091545, + 0.011933354660868645, + 0.007969394326210022, + 0.004009741824120283, + 0.00501700583845377, + 0.004483075346797705, + 0.0034657528158277273, + 0.003265716601163149, + 0.008345703594386578, + 0.0032563763670623302, + 0.010449101217091084, + 0.0016959871863946319, + 0.0012618614127859473, + 0.001403224770911038, + 0.0009297874639742076, + 0.0017866529524326324, + 0.0013441662304103374, + 0.001485921791754663, + 0.0011878038058057427, + 0.002144916681572795, + 0.0017502756090834737, + 0.0014570617349818349, + 0.0011894017225131392, + 0.0011939967516809702, + 0.0029681690502911806, + 0.0010392871918156743, + 0.0020243346225470304, + 0.0029104077257215977, + 0.004381491802632809, + 0.0026851268485188484, + 0.002781124319881201, + 0.003372799139469862, + 0.005747898016124964, + 0.0018347528530284762, + 0.004739115014672279, + 0.03281063213944435, + 0.03236428275704384, + 0.032501913607120514, + 0.01774444803595543, + 0.06291379779577255, + 0.05332356318831444, + 0.01552718784660101, + 0.026176046580076218, + 0.003767220303416252, + 0.0019869038369506598, + 0.0031674301717430353, + 0.003051956184208393, + 0.004023912828415632, + 0.0033833826892077923, + 0.0019265719456598163, + 0.005955555476248264, + 0.00808451697230339, + 0.00949487742036581, + 0.005718565080314875, + 0.004216715693473816, + 0.013165009208023548, + 0.012034263461828232, + 0.002981252036988735, + 0.0038625833112746477, + 0.008288650773465633, + 0.08837268501520157, + 0.003362948540598154, + 0.003268191358074546, + 0.09210087358951569, + 0.009237599559128284, + 0.0036014914512634277, + 0.004038854502141476, + 0.006151411682367325, + 0.11282560974359512, + 0.004102781880646944, + 0.00512683903798461, + 0.11245005577802658, + 0.008411327376961708, + 0.004364626947790384, + 0.0040804981254041195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.04303782060742378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal medium shot focuses on a scene composed entirely of pixels, embodying a retro 8-bit gaming aesthetic. A wooden bookshelf, composed of dark and light brown pixels, dominates the frame, neatly arranged with vibrantly colored books featuring distinct pixelated edges. Their spines display saturated red, emerald green, royal blue, and bright yellow. A simple, light-gray pixelated wall frames the background, casting an even, bright light across the frame, creating virtually no shadows. The action begins with a red book at the end of the queue cautiously tilting to one side before quickly losing balance and crashing into a nearby green book, triggering a chain reaction. Immediately, the books fall one by one in a slightly stuttering, frame-by-frame animation, tumbling and plummeting through the air before disappearing off the bottom of the frame. The entire sequence is imbued with a sense of dynamism and nostalgia. The vibrant colors and unique pixelated dynamics create a lighthearted, playful atmosphere.", + "seed": 28606, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_1438_seed_28606_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": -0.4427867688238621, + "vlm_logprob": -0.4427867688238621, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.4427867688238621, + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.3784729540348053 + ], + "opposite_logprob_tokens": [ + -1.253472924232483 + ], + "logprob_reward": -0.3784729540348053, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3784729540348053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1433068960905075, + 0.18093335628509521, + 0.04416137933731079, + 0.17287907004356384, + 0.04095463082194328, + 0.12356165796518326, + 0.08965665847063065, + 0.06705179065465927, + 0.060685910284519196, + 0.07680866122245789 + ], + "credit_map": [ + 0.0306104626506567, + 0.01118842326104641, + 0.014242468401789665, + 0.018581504002213478, + 0.027038192376494408, + 0.015366778708994389, + 0.012079783715307713, + 0.014199291355907917, + 0.051635656505823135, + 0.015649298205971718, + 0.021526314318180084, + 0.010908444412052631, + 0.05008767172694206, + 0.011413777247071266, + 0.007947600446641445, + 0.011764602735638618, + 0.004173389635980129, + 0.009881565347313881, + 0.003741111606359482, + 0.0036422403063625097, + 0.005151330027729273, + 0.00739189563319087, + 0.004036553669720888, + 0.006143294740468264, + 0.013676280155777931, + 0.0635158121585846, + 0.0075982194393873215, + 0.006316567771136761, + 0.010148212313652039, + 0.05700172111392021, + 0.0055004386231303215, + 0.009121817536652088, + 0.00600165780633688, + 0.007954302243888378, + 0.005016886163502932, + 0.0036738435737788677, + 0.004200293216854334, + 0.005759820807725191, + 0.003981797955930233, + 0.00436603045091033, + 0.011467677541077137, + 0.008980788290500641, + 0.012087649665772915, + 0.00501713203266263, + 0.04702033847570419, + 0.00970778614282608, + 0.007595169357955456, + 0.021685119718313217, + 0.01537376455962658, + 0.013750971294939518, + 0.008729345165193081, + 0.008812187239527702, + 0.01349416933953762, + 0.00985627993941307, + 0.009900795295834541, + 0.00973914097994566, + 0.007954937405884266, + 0.008646058849990368, + 0.013094155117869377, + 0.004121948499232531, + 0.013208876363933086, + 0.007419840432703495, + 0.006712953094393015, + 0.005893019959330559, + 0.004958975128829479, + 0.005492216907441616, + 0.010980933904647827, + 0.007573096547275782, + 0.0069479248486459255, + 0.007765049580484629, + 0.007032481487840414, + 0.00993522722274065, + 0.005535250995308161, + 0.006352802738547325, + 0.010964895598590374, + 0.010376987978816032, + 0.009210027754306793, + 0.01047453097999096, + 0.006586344912648201, + 0.017307821661233902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.3784729540348053 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.701539158821106 + ], + "opposite_logprob_tokens": [ + -0.20153914391994476 + ], + "logprob_reward": -1.701539158821106, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.701539158821106, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06935621052980423, + 0.2873188555240631, + 0.02636999636888504, + 0.07737192511558533, + 0.09462574869394302, + 0.12824635207653046, + 0.10493860393762589, + 0.06373920291662216, + 0.0902354046702385, + 0.05779765546321869 + ], + "credit_map": [ + 0.013803965412080288, + 0.005631219130009413, + 0.0062456438317894936, + 0.007073949556797743, + 0.012676955200731754, + 0.00933486595749855, + 0.006467938423156738, + 0.008121680468320847, + 0.07314053922891617, + 0.0295316893607378, + 0.009141303598880768, + 0.010535216890275478, + 0.12618213891983032, + 0.014077900908887386, + 0.01321671437472105, + 0.011493386700749397, + 0.0030539848376065493, + 0.004810857120901346, + 0.002288484713062644, + 0.002758623333647847, + 0.002797130960971117, + 0.002764705801382661, + 0.0033543703611940145, + 0.004541843663901091, + 0.014711429364979267, + 0.01009480468928814, + 0.006082931533455849, + 0.005440979730337858, + 0.006477931514382362, + 0.019921286031603813, + 0.00442078709602356, + 0.01022177841514349, + 0.017591116949915886, + 0.0180071834474802, + 0.010464558377861977, + 0.00776498019695282, + 0.010011749342083931, + 0.013961484655737877, + 0.009487262926995754, + 0.007337416056543589, + 0.026113437488675117, + 0.010319911874830723, + 0.009914830327033997, + 0.010149562731385231, + 0.04225868359208107, + 0.01291153859347105, + 0.008446147665381432, + 0.00813225470483303, + 0.026280811056494713, + 0.013697761110961437, + 0.008397598750889301, + 0.009050078690052032, + 0.019890297204256058, + 0.012938620522618294, + 0.0072120120748877525, + 0.007471439428627491, + 0.006720658857375383, + 0.008216261863708496, + 0.015133015811443329, + 0.005110493395477533, + 0.011697264388203621, + 0.006938113830983639, + 0.005659906659275293, + 0.004263496957719326, + 0.007127922028303146, + 0.010374183766543865, + 0.018831979483366013, + 0.009586517699062824, + 0.009655197151005268, + 0.018259068951010704, + 0.0076285358518362045, + 0.008772009052336216, + 0.0045547159388661385, + 0.00519648427143693, + 0.007112423423677683, + 0.0073202853091061115, + 0.005374251864850521, + 0.00722160842269659, + 0.0053895157761871815, + 0.015628375113010406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -1.701539158821106 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.009824376553297043 + ], + "opposite_logprob_tokens": [ + -4.634824275970459 + ], + "logprob_reward": -0.009824376553297043, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009824376553297043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051508430391550064, + 0.19953298568725586, + 0.038608260452747345, + 0.11226986348628998, + 0.09239928424358368, + 0.15561705827713013, + 0.1210298091173172, + 0.0940937027335167, + 0.07353793829679489, + 0.06140265613794327 + ], + "credit_map": [ + 0.008078405633568764, + 0.0053312890231609344, + 0.005786158610135317, + 0.0062658763490617275, + 0.00772101990878582, + 0.007805683184415102, + 0.005033695604652166, + 0.005486303940415382, + 0.049108028411865234, + 0.034746572375297546, + 0.014004696160554886, + 0.012494755908846855, + 0.018071001395583153, + 0.013427197001874447, + 0.02082838863134384, + 0.03685235232114792, + 0.007016520947217941, + 0.008160046301782131, + 0.0025773351080715656, + 0.003064855933189392, + 0.004244244657456875, + 0.0035900494549423456, + 0.004297587554901838, + 0.005657624453306198, + 0.022720249369740486, + 0.012178605422377586, + 0.010744454339146614, + 0.008080275729298592, + 0.01160490047186613, + 0.02187502197921276, + 0.006550632882863283, + 0.018515726551413536, + 0.01628170907497406, + 0.016116710379719734, + 0.009207282215356827, + 0.008361299522221088, + 0.010608595795929432, + 0.014316653832793236, + 0.009958388283848763, + 0.007548653054982424, + 0.036038968712091446, + 0.011663812212646008, + 0.012141321785748005, + 0.010565076023340225, + 0.05150798708200455, + 0.014994857832789421, + 0.010348175652325153, + 0.008356855250895023, + 0.026143940165638924, + 0.017300445586442947, + 0.009380470961332321, + 0.01150114182382822, + 0.0208735428750515, + 0.01660831831395626, + 0.009480934590101242, + 0.009741018526256084, + 0.01053711585700512, + 0.014614254236221313, + 0.021600570529699326, + 0.007878394797444344, + 0.017031699419021606, + 0.008607864379882812, + 0.00887968484312296, + 0.004944127518683672, + 0.006768125109374523, + 0.007573884911835194, + 0.012557903304696083, + 0.008896558545529842, + 0.008009721525013447, + 0.013672682456672192, + 0.008480465039610863, + 0.007578604388982058, + 0.00476820906624198, + 0.004994784947484732, + 0.009850879199802876, + 0.006715685129165649, + 0.0048764594830572605, + 0.007318629417568445, + 0.005872625857591629, + 0.01700538583099842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.009824376553297043 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.062054336071014404 + ], + "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": [ + 0.05122348293662071, + 0.2706715762615204, + 0.02352651208639145, + 0.08761553466320038, + 0.07245700806379318, + 0.07105345278978348, + 0.09200406819581985, + 0.11979980021715164, + 0.15182237327098846, + 0.05982617288827896 + ], + "credit_map": [ + 0.009524114429950714, + 0.005200332496315241, + 0.006783968303352594, + 0.0059010242111980915, + 0.005670627113431692, + 0.006739373784512281, + 0.004729037638753653, + 0.006674999836832285, + 0.071865975856781, + 0.025532767176628113, + 0.010611284524202347, + 0.010175132192671299, + 0.12025269865989685, + 0.014086730778217316, + 0.010451553389430046, + 0.007695429492741823, + 0.0032184983137995005, + 0.0032255342230200768, + 0.0021242653019726276, + 0.002006515860557556, + 0.0025733057409524918, + 0.0026450851000845432, + 0.0032098821830004454, + 0.0045234267599880695, + 0.014756315387785435, + 0.025827931240200996, + 0.005367865785956383, + 0.0051157246343791485, + 0.005719767417758703, + 0.015231091529130936, + 0.004438849166035652, + 0.01115798857063055, + 0.0123712457716465, + 0.013925991952419281, + 0.007467994000762701, + 0.006534358020871878, + 0.009208324365317822, + 0.009887210093438625, + 0.007113501895219088, + 0.005948380101472139, + 0.006956929340958595, + 0.004869930446147919, + 0.011047537438571453, + 0.0036078698467463255, + 0.021920204162597656, + 0.007909930311143398, + 0.005586438346654177, + 0.009154606610536575, + 0.016316670924425125, + 0.014351160265505314, + 0.007552686147391796, + 0.008853554725646973, + 0.01399722695350647, + 0.012864749878644943, + 0.008340262807905674, + 0.009727752767503262, + 0.01225725095719099, + 0.021539969369769096, + 0.026407670229673386, + 0.0090556088835001, + 0.024359913542866707, + 0.009810860268771648, + 0.010342283174395561, + 0.006026248447597027, + 0.01725226826965809, + 0.01645328849554062, + 0.028283480554819107, + 0.01887631043791771, + 0.014376861043274403, + 0.02191489189863205, + 0.015085326507687569, + 0.019579937681555748, + 0.0038669982459396124, + 0.005681727547198534, + 0.010927554219961166, + 0.005895067472010851, + 0.004989800043404102, + 0.00679659191519022, + 0.00550589757040143, + 0.016162527725100517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.06204301863908768 + ], + "opposite_logprob_tokens": [ + -2.8120429515838623 + ], + "logprob_reward": -0.06204301863908768, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204301863908768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062106162309646606, + 0.06491870433092117, + 0.04390672594308853, + 0.05056777969002724, + 0.04193075746297836, + 0.12850667536258698, + 0.08064713329076767, + 0.17947041988372803, + 0.2661711871623993, + 0.08177446573972702 + ], + "credit_map": [ + 0.010694693773984909, + 0.006639948580414057, + 0.009028526954352856, + 0.011473013088107109, + 0.010007859207689762, + 0.0055106692016124725, + 0.004446176812052727, + 0.0043052732944488525, + 0.018068162724375725, + 0.00702870637178421, + 0.004750343970954418, + 0.004398629534989595, + 0.010507756844162941, + 0.004716831725090742, + 0.00560583034530282, + 0.009842432104051113, + 0.008539756760001183, + 0.009261943399906158, + 0.004137647338211536, + 0.0029793372377753258, + 0.003309148596599698, + 0.005510910414159298, + 0.004618212580680847, + 0.005549765657633543, + 0.010137306526303291, + 0.006637848448008299, + 0.0043867202475667, + 0.0034801869187504053, + 0.0037906761281192303, + 0.010456804186105728, + 0.003093996085226536, + 0.008584236726164818, + 0.007963712327182293, + 0.007592900190502405, + 0.004834597930312157, + 0.004527969751507044, + 0.004102029837667942, + 0.0061795650981366634, + 0.003707634285092354, + 0.003022346179932356, + 0.014027997851371765, + 0.009583951905369759, + 0.030021661892533302, + 0.013660680502653122, + 0.030884334817528725, + 0.012309935875236988, + 0.010189123451709747, + 0.007828972302377224, + 0.012407452799379826, + 0.011377708986401558, + 0.0061513702385127544, + 0.007265482563525438, + 0.01246632356196642, + 0.009874069131910801, + 0.008044426329433918, + 0.01306029874831438, + 0.02001030370593071, + 0.03184675797820091, + 0.03777419030666351, + 0.012864318676292896, + 0.03670152649283409, + 0.014762690290808678, + 0.017131373286247253, + 0.008379252627491951, + 0.03198697045445442, + 0.024298537522554398, + 0.041763514280319214, + 0.03446931019425392, + 0.022330712527036667, + 0.038417715579271317, + 0.03194248676300049, + 0.04096192866563797, + 0.0062882076017558575, + 0.008321279659867287, + 0.011446070857346058, + 0.011110412888228893, + 0.008176340721547604, + 0.009086929261684418, + 0.006197463255375624, + 0.02114775963127613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.06204301863908768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, horizontal medium shot focuses on a scene composed entirely of pixels, embodying a retro 8-bit gaming aesthetic. A wooden bookshelf, composed of dark and light brown pixels, dominates the frame, neatly arranged with vibrantly colored books featuring distinct pixelated edges. Their spines display saturated red, emerald green, royal blue, and bright yellow. A simple, light-gray pixelated wall frames the background, casting an even, bright light across the frame, creating virtually no shadows. The action begins with a red book at the end of the queue cautiously tilting to one side before quickly losing balance and crashing into a nearby green book, triggering a chain reaction. Immediately, the books fall one by one in a slightly stuttering, frame-by-frame animation, tumbling and plummeting through the air before disappearing off the bottom of the frame. The entire sequence is imbued with a sense of dynamism and nostalgia. The vibrant colors and unique pixelated dynamics create a lighthearted, playful atmosphere.", + "seed": 28607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000017/group_9/index_1438_seed_28607_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": -0.06959704770706594, + "vlm_logprob": -0.06959704770706594, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.06959704770706594, + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.049456559121608734 + ], + "opposite_logprob_tokens": [ + -3.7994565963745117 + ], + "logprob_reward": -0.049456559121608734, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.049456559121608734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1357843279838562, + 0.12081767618656158, + 0.04658770188689232, + 0.05144529044628143, + 0.06068575009703636, + 0.1265488862991333, + 0.08185809850692749, + 0.18392768502235413, + 0.10162918269634247, + 0.09071539342403412 + ], + "credit_map": [ + 0.016133805736899376, + 0.011104785837233067, + 0.011754748411476612, + 0.01233182568103075, + 0.017563579604029655, + 0.021343573927879333, + 0.019377898424863815, + 0.026174113154411316, + 0.007612436078488827, + 0.010347853414714336, + 0.008231322281062603, + 0.009438863024115562, + 0.032497845590114594, + 0.021544797345995903, + 0.012768968008458614, + 0.01837560161948204, + 0.011278621852397919, + 0.00399783905595541, + 0.003590006148442626, + 0.004887328948825598, + 0.006443848367780447, + 0.007113955449312925, + 0.0035235423129051924, + 0.00575255835428834, + 0.008048833347856998, + 0.004731778521090746, + 0.004740605596452951, + 0.00406889570876956, + 0.009053818881511688, + 0.009842408820986748, + 0.004444299265742302, + 0.006514652632176876, + 0.008838463574647903, + 0.0054731168784201145, + 0.0051636891439557076, + 0.0047413683496415615, + 0.010556894354522228, + 0.015326078981161118, + 0.005379341542720795, + 0.005206794012337923, + 0.016892991960048676, + 0.019574826583266258, + 0.012254712171852589, + 0.014151843264698982, + 0.014835505746304989, + 0.021016141399741173, + 0.012538952752947807, + 0.015283913351595402, + 0.021440625190734863, + 0.011564407497644424, + 0.008832309395074844, + 0.0059782820753753185, + 0.008192860521376133, + 0.008336663246154785, + 0.006407537497580051, + 0.011105409823358059, + 0.019114136695861816, + 0.018954386934638023, + 0.011709483340382576, + 0.01890706457197666, + 0.014425271190702915, + 0.034915171563625336, + 0.017047390341758728, + 0.04885478317737579, + 0.02478974498808384, + 0.013286333531141281, + 0.021321730688214302, + 0.007639291696250439, + 0.005181369371712208, + 0.012619007378816605, + 0.006358377169817686, + 0.010433322750031948, + 0.014559590257704258, + 0.01774434745311737, + 0.01272329781204462, + 0.007693307474255562, + 0.0056566703133285046, + 0.01685292460024357, + 0.006285268347710371, + 0.009199986234307289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.049456559121608734 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.016091594472527504 + ], + "opposite_logprob_tokens": [ + -4.141091823577881 + ], + "logprob_reward": -0.016091594472527504, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091594472527504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07832114398479462, + 0.04225720837712288, + 0.09016081690788269, + 0.2543286383152008, + 0.052329469472169876, + 0.17817072570323944, + 0.07625960558652878, + 0.1067432165145874, + 0.05479719862341881, + 0.06663196533918381 + ], + "credit_map": [ + 0.008967419154942036, + 0.0061639221385121346, + 0.009168777614831924, + 0.011619864031672478, + 0.009955475106835365, + 0.011138781905174255, + 0.007981507107615471, + 0.013325396925210953, + 0.00518563250079751, + 0.004562589339911938, + 0.004594005178660154, + 0.0033839175011962652, + 0.01009589247405529, + 0.006730926223099232, + 0.003861848497763276, + 0.0038424008525907993, + 0.01584073342382908, + 0.006334630772471428, + 0.007894592359662056, + 0.008911392651498318, + 0.010334011167287827, + 0.012569827027618885, + 0.010113832540810108, + 0.018161794170737267, + 0.019656319171190262, + 0.0109977712854743, + 0.0076690842397511005, + 0.010762857273221016, + 0.08201975375413895, + 0.10573316365480423, + 0.0076174503192305565, + 0.00987224280834198, + 0.008428903296589851, + 0.006832103710621595, + 0.005983020644634962, + 0.006428409833461046, + 0.007433220278471708, + 0.005721768829971552, + 0.006812655366957188, + 0.004689384717494249, + 0.020657207816839218, + 0.027054332196712494, + 0.018678350374102592, + 0.020850416272878647, + 0.024280352517962456, + 0.02015346847474575, + 0.024873754009604454, + 0.021622849628329277, + 0.022150572389364243, + 0.009121716022491455, + 0.010564344935119152, + 0.006306091323494911, + 0.005940926261246204, + 0.008866618387401104, + 0.005026270169764757, + 0.00828306470066309, + 0.016456851735711098, + 0.014714333228766918, + 0.008098697289824486, + 0.010978835634887218, + 0.01017146185040474, + 0.02403242699801922, + 0.009051329456269741, + 0.013239272870123386, + 0.012874212116003036, + 0.006371656898409128, + 0.010691986419260502, + 0.004495079629123211, + 0.003034254303202033, + 0.006798722315579653, + 0.004581709858030081, + 0.0059495773166418076, + 0.010217183269560337, + 0.010997282341122627, + 0.010455437935888767, + 0.007186645641922951, + 0.004069545306265354, + 0.01230330765247345, + 0.005593244917690754, + 0.005809321999549866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.016091594472527504 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.0019550274591892958 + ], + "opposite_logprob_tokens": [ + -6.251955032348633 + ], + "logprob_reward": -0.0019550274591892958, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019550274591892958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.106196828186512, + 0.2150057703256607, + 0.08552596718072891, + 0.11906080693006516, + 0.09640204906463623, + 0.09008698165416718, + 0.10509079694747925, + 0.08859188109636307, + 0.03701375424861908, + 0.05702519044280052 + ], + "credit_map": [ + 0.012742441147565842, + 0.011116135865449905, + 0.009127159602940083, + 0.014470766298472881, + 0.016115574166178703, + 0.013568645343184471, + 0.016343487426638603, + 0.012712621130049229, + 0.020268170163035393, + 0.022869789972901344, + 0.021350577473640442, + 0.018281735479831696, + 0.054148200899362564, + 0.03662944585084915, + 0.023033546283841133, + 0.018424298614263535, + 0.009418982081115246, + 0.008659898303449154, + 0.007314292248338461, + 0.0073821996338665485, + 0.02330908179283142, + 0.015272202901542187, + 0.006227320060133934, + 0.007941985502839088, + 0.009099263697862625, + 0.0085606025531888, + 0.006246234755963087, + 0.00885623786598444, + 0.029675554484128952, + 0.038515854626894, + 0.007619787007570267, + 0.0104872677475214, + 0.011466619558632374, + 0.007468924392014742, + 0.013480538502335548, + 0.007664166856557131, + 0.01535301934927702, + 0.022871332243084908, + 0.008100086823105812, + 0.009997357614338398, + 0.011908415704965591, + 0.009163184091448784, + 0.007469892501831055, + 0.00998581200838089, + 0.007685561198741198, + 0.019961856305599213, + 0.011477704159915447, + 0.012434552423655987, + 0.01932479627430439, + 0.01295364648103714, + 0.024827688932418823, + 0.009457684122025967, + 0.008348312228918076, + 0.012283294461667538, + 0.008099693804979324, + 0.009795687161386013, + 0.015189549885690212, + 0.011363616213202477, + 0.010433495976030827, + 0.006988362409174442, + 0.008026759140193462, + 0.012548905797302723, + 0.006100584752857685, + 0.01794060505926609, + 0.00431361747905612, + 0.006162825971841812, + 0.0029864509124308825, + 0.005178955849260092, + 0.002890601521357894, + 0.005385436117649078, + 0.004309718031436205, + 0.0057861460372805595, + 0.009641465730965137, + 0.010874210856854916, + 0.006707057822495699, + 0.006103446241468191, + 0.004184461664408445, + 0.008668887428939342, + 0.004186645615845919, + 0.0066590141505002975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0019550274591892958 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.22554773092269897 + ], + "opposite_logprob_tokens": [ + -1.6005477905273438 + ], + "logprob_reward": -0.22554773092269897, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554773092269897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10394057631492615, + 0.18793228268623352, + 0.02290116436779499, + 0.10464764386415482, + 0.047201842069625854, + 0.12696561217308044, + 0.15328195691108704, + 0.10329872369766235, + 0.12075739353895187, + 0.029072849079966545 + ], + "credit_map": [ + 0.01547168381512165, + 0.0074270642362535, + 0.009567560628056526, + 0.011500582098960876, + 0.015053680166602135, + 0.013511096127331257, + 0.01719038560986519, + 0.014218509197235107, + 0.01458622794598341, + 0.015872079879045486, + 0.014060835354030132, + 0.013830676674842834, + 0.049516286700963974, + 0.03574927896261215, + 0.01908796839416027, + 0.025228913873434067, + 0.003565909108147025, + 0.0022662163246423006, + 0.0026742450427263975, + 0.001747815404087305, + 0.0031008548103272915, + 0.0037388051860034466, + 0.002293187193572521, + 0.0035141301341354847, + 0.01427948847413063, + 0.004550054203718901, + 0.005206947214901447, + 0.004954742267727852, + 0.02559923380613327, + 0.03286478668451309, + 0.005786411929875612, + 0.011405978351831436, + 0.007213049102574587, + 0.005496698897331953, + 0.004803644027560949, + 0.0053029740229249, + 0.006530012935400009, + 0.005891347303986549, + 0.00732366181910038, + 0.004640454892069101, + 0.017644289880990982, + 0.019212907180190086, + 0.012191040441393852, + 0.015180408023297787, + 0.012134660966694355, + 0.022559121251106262, + 0.013465815223753452, + 0.014577355235815048, + 0.03353304788470268, + 0.02161780558526516, + 0.03397979959845543, + 0.016373882070183754, + 0.009674407541751862, + 0.015539075247943401, + 0.011604635044932365, + 0.010959304869174957, + 0.020151158794760704, + 0.012134932912886143, + 0.01304912194609642, + 0.007909563370049, + 0.013144905678927898, + 0.014023126102983952, + 0.008510353974997997, + 0.014375552535057068, + 0.03051868826150894, + 0.016520662233233452, + 0.02484942227602005, + 0.008702057413756847, + 0.006365274544805288, + 0.015059399418532848, + 0.007982902228832245, + 0.010758980177342892, + 0.0035474556498229504, + 0.0026797805912792683, + 0.004427976906299591, + 0.005556813906878233, + 0.0020906683057546616, + 0.0027010918129235506, + 0.0024070357903838158, + 0.005662024021148682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.22554773092269897 + }, + { + "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 visual style composed entirely of pixels with a retro 8-bit gaming aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a wooden bookshelf with dark and light brown pixels dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the books display saturated red, emerald green, royal blue, and bright yellow colors with pixelated edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a red book tilt, lose balance, and crash into a green book to trigger a chain reaction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the books fall one by one in a stuttering, frame-by-frame animation before disappearing off the bottom of the 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": [ + -0.05493432655930519 + ], + "opposite_logprob_tokens": [ + -2.92993426322937 + ], + "logprob_reward": -0.05493432655930519, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493432655930519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08551499992609024, + 0.06837594509124756, + 0.044839464128017426, + 0.059973251074552536, + 0.09926324337720871, + 0.35419660806655884, + 0.05150690674781799, + 0.10409243404865265, + 0.05624670535326004, + 0.07599041610956192 + ], + "credit_map": [ + 0.011171115562319756, + 0.006234857253730297, + 0.006991151254624128, + 0.009123558178544044, + 0.010493626818060875, + 0.011346041224896908, + 0.008996381424367428, + 0.021158263087272644, + 0.00451644416898489, + 0.0051996721886098385, + 0.0033959639258682728, + 0.0038564701098948717, + 0.021012676879763603, + 0.012351421639323235, + 0.00808483362197876, + 0.009958464652299881, + 0.009621472097933292, + 0.004049800336360931, + 0.004556517116725445, + 0.0035573572386056185, + 0.006874431390315294, + 0.00589254405349493, + 0.004009189084172249, + 0.006278154440224171, + 0.008887811563909054, + 0.005601623561233282, + 0.004978373181074858, + 0.004230825696140528, + 0.010201042518019676, + 0.012678408995270729, + 0.004551212303340435, + 0.008843950927257538, + 0.01530841551721096, + 0.012413469143211842, + 0.010195007547736168, + 0.011007454246282578, + 0.012371146120131016, + 0.012659929692745209, + 0.014484847895801067, + 0.01082297507673502, + 0.04857705160975456, + 0.05527443438768387, + 0.03386995941400528, + 0.0416865237057209, + 0.04260479658842087, + 0.0537121556699276, + 0.033145301043987274, + 0.045326363295316696, + 0.011391149833798409, + 0.00700036808848381, + 0.005970431957393885, + 0.004093917086720467, + 0.005423448514193296, + 0.0081132547929883, + 0.004001700785011053, + 0.005512636620551348, + 0.014778731390833855, + 0.01048178318887949, + 0.00889501627534628, + 0.009387202560901642, + 0.007616549730300903, + 0.013705252669751644, + 0.007221919484436512, + 0.03200598061084747, + 0.007133974693715572, + 0.00947590358555317, + 0.007263380568474531, + 0.007921998389065266, + 0.0038297902792692184, + 0.007758354768157005, + 0.005586477927863598, + 0.00727682514116168, + 0.009285585954785347, + 0.01542935986071825, + 0.009309954941272736, + 0.0071867164224386215, + 0.006201964803040028, + 0.014226231724023819, + 0.0058337426744401455, + 0.00851686391979456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.05493432655930519 + } + ], + "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/0000060/group_0/index_tensor([7288, 7288])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_tensor([7288, 7288])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d436a160e5b1a22c29f66280ff11eaeecd65d6e9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_tensor([7288, 7288])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot, clinging just inches above the worn cobblestones, fluidly weaves between two figures engaged in a high-speed roller-skating chase across a vibrant, crowded open-air market. The camera begins by following a young woman, her long brown hair tied in a high ponytail, flowing in the air as she glances back with a smile of excitement on her face. Dressed in a bright yellow tank top and denim shorts, she expertly navigates her scuffed burgundy roller skates through the market stalls, her movements swift and precise. The camera quickly pans to capture her pursuer, a young man with wind-torn shorts and a low, playful determination. Dressed in a royal blue windbreaker and black gym shorts, he propels his black and silver skates forward with a powerful thrust across the uneven ground. They race in a sensual explosion of color and movement: stalls overflowing with red tomatoes and green vegetables, awnings draped with vibrant textiles, and the diverse crowd of shoppers reduced to a soft, impressionistic blur. The bright midday sun illuminates the scene, filtering through the gaps between the stalls and casting fast-moving, speckled patterns of light and shadow on the ground, creating an atmosphere of energy, joy, vitality, and youthful freedom.", + "seed": 20155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_7288_seed_20155_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": -0.18952582646161317, + "vlm_logprob": -0.18952582646161317, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.18952582646161317, + "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 taken from a low angle, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8798960447311401 + ], + "opposite_logprob_tokens": [ + -0.6298960447311401 + ], + "logprob_reward": -0.8798960447311401, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8798960447311401, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3438819944858551, + 0.09099999815225601, + 0.0938001498579979, + 0.06419217586517334, + 0.03183837980031967, + 0.16180165112018585, + 0.06685822457075119, + 0.0323517844080925, + 0.0474904365837574, + 0.06678526848554611 + ], + "credit_map": [ + 0.061639245599508286, + 0.038381773978471756, + 0.06238296627998352, + 0.04522183910012245, + 0.03863852098584175, + 0.03275079280138016, + 0.026319697499275208, + 0.03854713961482048, + 0.01258463691920042, + 0.01043581310659647, + 0.00973589438945055, + 0.011486136354506016, + 0.013164895586669445, + 0.007347763050347567, + 0.01599201001226902, + 0.01025285106152296, + 0.012568037025630474, + 0.01016785018146038, + 0.009103942662477493, + 0.013683533295989037, + 0.009981639683246613, + 0.008598152548074722, + 0.017186766490340233, + 0.012510226108133793, + 0.011653685942292213, + 0.008989103138446808, + 0.005441033747047186, + 0.009145003743469715, + 0.005772531032562256, + 0.00613531656563282, + 0.007236305624246597, + 0.009819199331104755, + 0.0043487283401191235, + 0.003993578255176544, + 0.003862515790387988, + 0.0045233918353915215, + 0.00484446482732892, + 0.002541585126891732, + 0.004278963897377253, + 0.003445153823122382, + 0.02104106917977333, + 0.030230863019824028, + 0.025613535195589066, + 0.021434538066387177, + 0.014478256925940514, + 0.019322777166962624, + 0.015031296759843826, + 0.014649310149252415, + 0.008299585431814194, + 0.006335605401545763, + 0.008960430510342121, + 0.013883006758987904, + 0.009114489890635014, + 0.004722175654023886, + 0.008285092189908028, + 0.007257838733494282, + 0.003241033060476184, + 0.0033446720335632563, + 0.004651010502129793, + 0.006629692856222391, + 0.0038814719300717115, + 0.0036158664152026176, + 0.004165971651673317, + 0.0028220685198903084, + 0.003332034219056368, + 0.003969280980527401, + 0.006998436991125345, + 0.008826863951981068, + 0.007246571127325296, + 0.0038408825639635324, + 0.009047308005392551, + 0.004229057114571333, + 0.008631998673081398, + 0.00611490523442626, + 0.00880327820777893, + 0.009028090164065361, + 0.013640922494232655, + 0.0065919398330152035, + 0.007167269941419363, + 0.0068068658001720905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.8798960447311401 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009818828664720058 + ], + "opposite_logprob_tokens": [ + -4.634819030761719 + ], + "logprob_reward": -0.009818828664720058, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009818828664720058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34028875827789307, + 0.06626075506210327, + 0.09116753935813904, + 0.05827431380748749, + 0.0461595244705677, + 0.2620696723461151, + 0.027425434440374374, + 0.03260757029056549, + 0.03804633021354675, + 0.03770001232624054 + ], + "credit_map": [ + 0.0715503990650177, + 0.05111692473292351, + 0.06334348022937775, + 0.046240922063589096, + 0.02970992960035801, + 0.02286926843225956, + 0.02222215197980404, + 0.03323570266366005, + 0.009692990221083164, + 0.007254625670611858, + 0.010756009258329868, + 0.009510209783911705, + 0.008439538069069386, + 0.004570916760712862, + 0.007011044770479202, + 0.009025422856211662, + 0.019306667149066925, + 0.01107234787195921, + 0.009004991501569748, + 0.007717500440776348, + 0.017045937478542328, + 0.008687966503202915, + 0.010686325840651989, + 0.007645797915756702, + 0.007384606637060642, + 0.020782750099897385, + 0.011535242199897766, + 0.005461402703076601, + 0.003288296749815345, + 0.0028143455274403095, + 0.0044939094223082066, + 0.002513757674023509, + 0.005365984980016947, + 0.006096831988543272, + 0.005569302011281252, + 0.010421764105558395, + 0.00731302797794342, + 0.0036177770234644413, + 0.005136379972100258, + 0.0026384589727967978, + 0.021108556538820267, + 0.02790001779794693, + 0.06936293095350266, + 0.025395981967449188, + 0.013712857849895954, + 0.08514194190502167, + 0.008640974760055542, + 0.010806410573422909, + 0.0027107163332402706, + 0.0033113749232143164, + 0.004141935613006353, + 0.0059104491956532, + 0.0033299594651907682, + 0.002456222428008914, + 0.002651094924658537, + 0.0029136803932487965, + 0.0043732826597988605, + 0.0038808886893093586, + 0.004944407381117344, + 0.005083292257040739, + 0.002879777457565069, + 0.004934682976454496, + 0.0032244189642369747, + 0.0032868210691958666, + 0.009548074565827847, + 0.005834141746163368, + 0.005434209480881691, + 0.004132038913667202, + 0.003942662850022316, + 0.002829834120348096, + 0.0035041391383856535, + 0.0028212289325892925, + 0.008507505059242249, + 0.005846331361681223, + 0.005152799189090729, + 0.0060353754088282585, + 0.003757613943889737, + 0.0024988281074911356, + 0.0025173411704599857, + 0.003384219715371728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.009818828664720058 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645270295441151 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645270295441151, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645270295441151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15496157109737396, + 0.26464489102363586, + 0.023587340489029884, + 0.38818520307540894, + 0.026943780481815338, + 0.03850231692194939, + 0.029304541647434235, + 0.0208892822265625, + 0.018582085147500038, + 0.03439893573522568 + ], + "credit_map": [ + 0.033670879900455475, + 0.022634951397776604, + 0.013637198135256767, + 0.02248825505375862, + 0.019487129524350166, + 0.013739576563239098, + 0.01306240726262331, + 0.01624118722975254, + 0.025764301419258118, + 0.03449162840843201, + 0.022253913804888725, + 0.037968263030052185, + 0.04330819100141525, + 0.017283571884036064, + 0.0596955344080925, + 0.023879488930106163, + 0.0034079132601618767, + 0.0027993908151984215, + 0.0026879096403717995, + 0.0019285556627437472, + 0.0037203223910182714, + 0.004159936681389809, + 0.0025342535227537155, + 0.002349059097468853, + 0.03753555566072464, + 0.15631935000419617, + 0.08258114755153656, + 0.034265078604221344, + 0.020650073885917664, + 0.012759077362716198, + 0.030246064066886902, + 0.013828879222273827, + 0.004066244699060917, + 0.00325634959153831, + 0.004501400049775839, + 0.0038188283797353506, + 0.00449555367231369, + 0.001979488180950284, + 0.003002921584993601, + 0.001822993392124772, + 0.004507944919168949, + 0.005611715372651815, + 0.006037329323589802, + 0.004187033511698246, + 0.004226198419928551, + 0.00625417148694396, + 0.004177544731646776, + 0.003500379156321287, + 0.004332208074629307, + 0.0032622392755001783, + 0.0038782870396971703, + 0.0043113622814416885, + 0.0038346992805600166, + 0.0021537160500884056, + 0.004182575270533562, + 0.0033494560047984123, + 0.0022751588840037584, + 0.0020242929458618164, + 0.0026994487270712852, + 0.002968583256006241, + 0.0031034338753670454, + 0.003143757814541459, + 0.00287437392398715, + 0.0018002331489697099, + 0.002210996113717556, + 0.0024248429108411074, + 0.002711065113544464, + 0.0026377399917691946, + 0.0034045400097966194, + 0.0014485333813354373, + 0.0023619758430868387, + 0.0013823921326547861, + 0.0060589732602238655, + 0.004787761252373457, + 0.005100060719996691, + 0.006280885078012943, + 0.0033046577591449022, + 0.002626999979838729, + 0.0029328102245926857, + 0.0033067851327359676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.004645270295441151 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633286036550999 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633286036550999, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633286036550999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10896355658769608, + 0.22323434054851532, + 0.06703782826662064, + 0.06797897070646286, + 0.09833773225545883, + 0.21523378789424896, + 0.030539918690919876, + 0.06755717098712921, + 0.03351088985800743, + 0.0876058042049408 + ], + "credit_map": [ + 0.020923981443047523, + 0.01056739967316389, + 0.01666007749736309, + 0.017756152898073196, + 0.010311352089047432, + 0.01065930724143982, + 0.009762301109731197, + 0.012322983704507351, + 0.020216284319758415, + 0.027493739500641823, + 0.01825626939535141, + 0.02291199006140232, + 0.03647835925221443, + 0.01797669380903244, + 0.05283810570836067, + 0.027062904089689255, + 0.007888671942055225, + 0.00875631719827652, + 0.005957668647170067, + 0.008618120104074478, + 0.00836868491023779, + 0.005545653868466616, + 0.013746781274676323, + 0.00815592985600233, + 0.007597867399454117, + 0.015466516837477684, + 0.009252932853996754, + 0.007122305687516928, + 0.004669772461056709, + 0.007764753885567188, + 0.00654443446546793, + 0.009560382924973965, + 0.007817400619387627, + 0.009842860512435436, + 0.008916517719626427, + 0.015516843646764755, + 0.021556973457336426, + 0.008436399511992931, + 0.017977237701416016, + 0.008273500949144363, + 0.015269159339368343, + 0.021236198022961617, + 0.052089300006628036, + 0.019966119900345802, + 0.014350191690027714, + 0.07083998620510101, + 0.009447373449802399, + 0.012035461142659187, + 0.0028391468804329634, + 0.005083717405796051, + 0.0030444941949099302, + 0.003703334601595998, + 0.0044289217330515385, + 0.0029291450046002865, + 0.005740867927670479, + 0.0027702918741852045, + 0.006071616895496845, + 0.004760692827403545, + 0.011137085035443306, + 0.012552617117762566, + 0.008236471563577652, + 0.007918018847703934, + 0.010857679881155491, + 0.006022992543876171, + 0.002442748751491308, + 0.0025580015499144793, + 0.005204716697335243, + 0.005374541971832514, + 0.0067393118515610695, + 0.0030765864066779613, + 0.005338267423212528, + 0.0027767179999500513, + 0.011390550062060356, + 0.008117570541799068, + 0.012482243590056896, + 0.015566305257380009, + 0.01309630461037159, + 0.009006337262690067, + 0.010112551972270012, + 0.007833938114345074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.004633286036550999 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863570258021355 + ], + "opposite_logprob_tokens": [ + -3.048635721206665 + ], + "logprob_reward": -0.04863570258021355, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863570258021355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13406997919082642, + 0.06598030030727386, + 0.0967964306473732, + 0.18470299243927002, + 0.058060172945261, + 0.3271336853504181, + 0.03368717432022095, + 0.02373095601797104, + 0.02887849509716034, + 0.046959832310676575 + ], + "credit_map": [ + 0.02459648810327053, + 0.012644392438232899, + 0.014432218857109547, + 0.01485991757363081, + 0.013828701339662075, + 0.014045200310647488, + 0.017443174496293068, + 0.022219879552721977, + 0.0057548051699995995, + 0.0086703822016716, + 0.005264674313366413, + 0.010444552637636662, + 0.010433475486934185, + 0.004766344558447599, + 0.013560477644205093, + 0.0070855868980288506, + 0.01644711196422577, + 0.012323924340307713, + 0.009510836564004421, + 0.008075324818491936, + 0.020573128014802933, + 0.010779181495308876, + 0.010741213336586952, + 0.00834571197628975, + 0.020810702815651894, + 0.06634599715471268, + 0.0366300567984581, + 0.01782621070742607, + 0.010178887285292149, + 0.008188662119209766, + 0.013791645877063274, + 0.010930816642940044, + 0.008171706460416317, + 0.0064141275361180305, + 0.0088508240878582, + 0.01258687674999237, + 0.006441257428377867, + 0.004269254393875599, + 0.007128017954528332, + 0.004198107868432999, + 0.023497629910707474, + 0.03298543766140938, + 0.08036458492279053, + 0.02945578284561634, + 0.020316775888204575, + 0.11469430476427078, + 0.012573822401463985, + 0.013245326466858387, + 0.002400392899289727, + 0.0033278584014624357, + 0.005266419146209955, + 0.006917004007846117, + 0.004962450359016657, + 0.0024381745606660843, + 0.0044870018027722836, + 0.003887874772772193, + 0.0024836163502186537, + 0.00280868005938828, + 0.003492840565741062, + 0.004356028512120247, + 0.0024229225236922503, + 0.0033966812770813704, + 0.002575116930529475, + 0.002195068635046482, + 0.0034583937376737595, + 0.002840542234480381, + 0.00432949885725975, + 0.003678268287330866, + 0.004316034261137247, + 0.002747415564954281, + 0.0043468293733894825, + 0.003161514876410365, + 0.0062853326089680195, + 0.004831127356737852, + 0.006640045437961817, + 0.006744816899299622, + 0.006085681729018688, + 0.005462685599923134, + 0.00444073136895895, + 0.006469411309808493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.04863570258021355 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, clinging just inches above the worn cobblestones, fluidly weaves between two figures engaged in a high-speed roller-skating chase across a vibrant, crowded open-air market. The camera begins by following a young woman, her long brown hair tied in a high ponytail, flowing in the air as she glances back with a smile of excitement on her face. Dressed in a bright yellow tank top and denim shorts, she expertly navigates her scuffed burgundy roller skates through the market stalls, her movements swift and precise. The camera quickly pans to capture her pursuer, a young man with wind-torn shorts and a low, playful determination. Dressed in a royal blue windbreaker and black gym shorts, he propels his black and silver skates forward with a powerful thrust across the uneven ground. They race in a sensual explosion of color and movement: stalls overflowing with red tomatoes and green vegetables, awnings draped with vibrant textiles, and the diverse crowd of shoppers reduced to a soft, impressionistic blur. The bright midday sun illuminates the scene, filtering through the gaps between the stalls and casting fast-moving, speckled patterns of light and shadow on the ground, creating an atmosphere of energy, joy, vitality, and youthful freedom.", + "seed": 20156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_7288_seed_20156_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": -0.08128310088068247, + "vlm_logprob": -0.08128310088068247, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08128310088068247, + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3504328429698944 + ], + "opposite_logprob_tokens": [ + -1.4754328727722168 + ], + "logprob_reward": -0.3504328429698944, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3504328429698944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3563138246536255, + 0.06008794531226158, + 0.09278640896081924, + 0.040391791611909866, + 0.05925554409623146, + 0.09476941078901291, + 0.060992125421762466, + 0.06070312112569809, + 0.05093463137745857, + 0.12376516312360764 + ], + "credit_map": [ + 0.04569796472787857, + 0.04662678763270378, + 0.055044110864400864, + 0.030712619423866272, + 0.01939692720770836, + 0.031623367220163345, + 0.041581396013498306, + 0.08563067018985748, + 0.010821331292390823, + 0.006225245539098978, + 0.00829114206135273, + 0.007975194603204727, + 0.0059998976066708565, + 0.004283479414880276, + 0.009834330528974533, + 0.00665732566267252, + 0.013443445786833763, + 0.010970056056976318, + 0.01544662844389677, + 0.010153344832360744, + 0.006898215506225824, + 0.018535492941737175, + 0.009719288907945156, + 0.007619938813149929, + 0.003001053351908922, + 0.004426697269082069, + 0.0047527807764709, + 0.004975426476448774, + 0.00863571371883154, + 0.007255050353705883, + 0.003584798891097307, + 0.0037602693773806095, + 0.003776978235691786, + 0.007835669443011284, + 0.008224617689847946, + 0.007761192973703146, + 0.008211545646190643, + 0.007625871803611517, + 0.010355769656598568, + 0.005463895853608847, + 0.01116720587015152, + 0.013924404047429562, + 0.010160979814827442, + 0.007828728295862675, + 0.0140868304297328, + 0.012577710673213005, + 0.014111519791185856, + 0.010912035591900349, + 0.0054319375194609165, + 0.0053943125531077385, + 0.006289600860327482, + 0.00977008230984211, + 0.006055883131921291, + 0.009356636554002762, + 0.01033752504736185, + 0.008356149308383465, + 0.006719464436173439, + 0.004579924046993256, + 0.011056420393288136, + 0.007281030062586069, + 0.004965909756720066, + 0.010162916034460068, + 0.009119823575019836, + 0.006817634683102369, + 0.0073120431043207645, + 0.005593995563685894, + 0.005836424417793751, + 0.005590048618614674, + 0.0060118939727544785, + 0.00723954988643527, + 0.007794021628797054, + 0.005556656513363123, + 0.014215081930160522, + 0.019538814201951027, + 0.016502292826771736, + 0.012372978962957859, + 0.016437938436865807, + 0.013813273049890995, + 0.014437902718782425, + 0.016446884721517563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.3504328429698944 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016124792397022247 + ], + "opposite_logprob_tokens": [ + -4.141124725341797 + ], + "logprob_reward": -0.016124792397022247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016124792397022247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35777267813682556, + 0.03206903487443924, + 0.07654465734958649, + 0.044058363884687424, + 0.08922336995601654, + 0.07241394370794296, + 0.08078289777040482, + 0.14270208775997162, + 0.038927145302295685, + 0.06550583988428116 + ], + "credit_map": [ + 0.02996145933866501, + 0.0768304392695427, + 0.04883941635489464, + 0.11582182347774506, + 0.013020699843764305, + 0.01627713441848755, + 0.024600330740213394, + 0.032421376556158066, + 0.005927528254687786, + 0.0025842078030109406, + 0.0048672351986169815, + 0.003951541148126125, + 0.0034427056089043617, + 0.0025304669979959726, + 0.004572916775941849, + 0.004192432388663292, + 0.012011336162686348, + 0.005758264567703009, + 0.012593621388077736, + 0.006667819339782, + 0.007701379247009754, + 0.015283940359950066, + 0.008627139031887054, + 0.007901161909103394, + 0.0038775200955569744, + 0.0038419270422309637, + 0.008905858732759953, + 0.004011287819594145, + 0.005815796088427305, + 0.006274454295635223, + 0.0036465937737375498, + 0.007684925105422735, + 0.004706709645688534, + 0.009379244409501553, + 0.015110297128558159, + 0.01022057980298996, + 0.015754839405417442, + 0.013894958421587944, + 0.013845919631421566, + 0.006310818251222372, + 0.015631914138793945, + 0.012082085013389587, + 0.007441072259098291, + 0.008068324066698551, + 0.008636650629341602, + 0.007747186813503504, + 0.00551323127001524, + 0.007293472066521645, + 0.0065116663463413715, + 0.0046074180863797665, + 0.01923995092511177, + 0.022479118779301643, + 0.005338303744792938, + 0.008847820572555065, + 0.007579157594591379, + 0.006179459858685732, + 0.030561037361621857, + 0.012014334090054035, + 0.023401280865073204, + 0.01605239138007164, + 0.01276582945138216, + 0.015213397331535816, + 0.013507170602679253, + 0.019186651334166527, + 0.005073415581136942, + 0.004628836177289486, + 0.008078968152403831, + 0.0037622214294970036, + 0.0043152510188519955, + 0.0035389286931604147, + 0.0044721090234816074, + 0.005057414062321186, + 0.009836820885539055, + 0.0117955906316638, + 0.008064118213951588, + 0.011131330393254757, + 0.005513196811079979, + 0.0050476486794650555, + 0.007244249805808067, + 0.006872882600873709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.016124792397022247 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006750398315489292 + ], + "opposite_logprob_tokens": [ + -5.006750583648682 + ], + "logprob_reward": -0.006750398315489292, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006750398315489292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1300470381975174, + 0.10272610932588577, + 0.048764485865831375, + 0.22219665348529816, + 0.03892947733402252, + 0.07576832920312881, + 0.09220954775810242, + 0.05971476063132286, + 0.04255877807736397, + 0.18708473443984985 + ], + "credit_map": [ + 0.017869271337985992, + 0.031555384397506714, + 0.009023524820804596, + 0.014903790317475796, + 0.008749367669224739, + 0.01073505450040102, + 0.013432621955871582, + 0.02377803437411785, + 0.012653637677431107, + 0.009080164134502411, + 0.012113471515476704, + 0.01809856854379177, + 0.01646282710134983, + 0.007406855467706919, + 0.017326349392533302, + 0.009584244340658188, + 0.006449829787015915, + 0.004214657004922628, + 0.00771975377574563, + 0.005274022929370403, + 0.00753968209028244, + 0.008907578885555267, + 0.004717101808637381, + 0.003941870294511318, + 0.017313871532678604, + 0.019841298460960388, + 0.03128736466169357, + 0.022368140518665314, + 0.044575173407793045, + 0.037812162190675735, + 0.019754286855459213, + 0.029244398698210716, + 0.003335126442834735, + 0.0037943017669022083, + 0.004798033740371466, + 0.009027481079101562, + 0.003722819732502103, + 0.003845763625577092, + 0.006882612127810717, + 0.0035233465023338795, + 0.006181668024510145, + 0.008336583152413368, + 0.007661816198378801, + 0.010098394006490707, + 0.008432772010564804, + 0.011093458160758018, + 0.014707468450069427, + 0.009256179444491863, + 0.004918621852993965, + 0.005070229526609182, + 0.025465724989771843, + 0.018926817923784256, + 0.009199654683470726, + 0.008079136721789837, + 0.011385578662157059, + 0.009163803420960903, + 0.008604519069194794, + 0.006940898485481739, + 0.008534146472811699, + 0.006847892422229052, + 0.0072473157197237015, + 0.007283528335392475, + 0.00857099425047636, + 0.005685480777174234, + 0.005275821313261986, + 0.00506895687431097, + 0.005287897307425737, + 0.004186716396361589, + 0.0046847970224916935, + 0.0058349734172225, + 0.005742780398577452, + 0.0064768423326313496, + 0.020305469632148743, + 0.025548776611685753, + 0.023927366361021996, + 0.028598954901099205, + 0.029491955414414406, + 0.014545200392603874, + 0.02413574419915676, + 0.0205313041806221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.006750398315489292 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531585991382599 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531585991382599, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531585991382599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1570468246936798, + 0.1040378212928772, + 0.05690108984708786, + 0.04498264193534851, + 0.06415245682001114, + 0.03286929056048393, + 0.07553789764642715, + 0.05558883771300316, + 0.07298953086137772, + 0.3358936011791229 + ], + "credit_map": [ + 0.02075614593923092, + 0.010897275060415268, + 0.020127160474658012, + 0.061011902987957, + 0.00974360853433609, + 0.007785583846271038, + 0.013059701770544052, + 0.0136654581874609, + 0.010801377706229687, + 0.009177550673484802, + 0.011667558923363686, + 0.01728556677699089, + 0.018632525578141212, + 0.010124250315129757, + 0.014503597281873226, + 0.011845403350889683, + 0.006584594491869211, + 0.0053678592666983604, + 0.008878660388290882, + 0.005868170410394669, + 0.009291313588619232, + 0.010026520118117332, + 0.00593605637550354, + 0.004947920795530081, + 0.004461866803467274, + 0.006351062096655369, + 0.00518988398835063, + 0.00476721441373229, + 0.007422477472573519, + 0.00709901750087738, + 0.004779349081218243, + 0.004911777563393116, + 0.004499970003962517, + 0.007111222017556429, + 0.007941191084682941, + 0.010704424232244492, + 0.0065431720577180386, + 0.008631969802081585, + 0.011924531310796738, + 0.006795978639274836, + 0.00233443733304739, + 0.00398279819637537, + 0.0025763395242393017, + 0.0038795892614871264, + 0.004176766611635685, + 0.004576442763209343, + 0.0060476744547486305, + 0.005295244511216879, + 0.0043373750522732735, + 0.005002894904464483, + 0.017895538359880447, + 0.010560372844338417, + 0.008733079768717289, + 0.007614377420395613, + 0.01088469848036766, + 0.010509566403925419, + 0.007145372685045004, + 0.004097701981663704, + 0.010118527337908745, + 0.005393004510551691, + 0.005820902064442635, + 0.00853206217288971, + 0.005979518406093121, + 0.008501751348376274, + 0.009914745576679707, + 0.00884210504591465, + 0.00926453061401844, + 0.006833282765001059, + 0.008099410682916641, + 0.009538017213344574, + 0.01054183766245842, + 0.009955604560673237, + 0.036752138286828995, + 0.043490633368492126, + 0.033882495015859604, + 0.03918900340795517, + 0.05982382968068123, + 0.03767242282629013, + 0.0431232787668705, + 0.041959818452596664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012531585991382599 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575884729623795 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.020575884729623795, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575884729623795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34655848145484924, + 0.061597470194101334, + 0.08391281217336655, + 0.050180040299892426, + 0.09324724972248077, + 0.03725387528538704, + 0.06703273206949234, + 0.0518249087035656, + 0.04636421054601669, + 0.16202819347381592 + ], + "credit_map": [ + 0.02733735740184784, + 0.09236087650060654, + 0.042939458042383194, + 0.11065750569105148, + 0.014050651341676712, + 0.012319890782237053, + 0.02330907993018627, + 0.023583702743053436, + 0.009820439852774143, + 0.007361756172031164, + 0.008632331155240536, + 0.007237765472382307, + 0.008457202464342117, + 0.006029773969203234, + 0.0065001449547708035, + 0.00755805941298604, + 0.015382161363959312, + 0.008572936989367008, + 0.012172830291092396, + 0.009876439347863197, + 0.007395382970571518, + 0.013351056724786758, + 0.00857599824666977, + 0.008586014620959759, + 0.006838052999228239, + 0.005602172575891018, + 0.007087160833179951, + 0.004178645089268684, + 0.009079393930733204, + 0.008964184671640396, + 0.0039237067103385925, + 0.004506728146225214, + 0.00841091014444828, + 0.011090229265391827, + 0.014267480000853539, + 0.012198255397379398, + 0.012357083149254322, + 0.010864255018532276, + 0.015392999164760113, + 0.008666048757731915, + 0.003357586683705449, + 0.006438584066927433, + 0.003553935559466481, + 0.004307081922888756, + 0.005842151585966349, + 0.004318532533943653, + 0.005153743550181389, + 0.004282260313630104, + 0.0034185799304395914, + 0.004081022925674915, + 0.01743301749229431, + 0.01262897439301014, + 0.008267358876764774, + 0.005928859580308199, + 0.008083928376436234, + 0.007190990261733532, + 0.008295387029647827, + 0.006658855360001326, + 0.006813241168856621, + 0.007411717902868986, + 0.006032667588442564, + 0.005055164452642202, + 0.007021897938102484, + 0.004535982385277748, + 0.006089192349463701, + 0.005465073511004448, + 0.006145955994725227, + 0.004378084558993578, + 0.004796082153916359, + 0.005786010529845953, + 0.006433951202780008, + 0.007269863039255142, + 0.01909536123275757, + 0.015838073566555977, + 0.01791907474398613, + 0.018315516412258148, + 0.022366931661963463, + 0.025288250297307968, + 0.024371564388275146, + 0.018833432346582413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.020575884729623795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, clinging just inches above the worn cobblestones, fluidly weaves between two figures engaged in a high-speed roller-skating chase across a vibrant, crowded open-air market. The camera begins by following a young woman, her long brown hair tied in a high ponytail, flowing in the air as she glances back with a smile of excitement on her face. Dressed in a bright yellow tank top and denim shorts, she expertly navigates her scuffed burgundy roller skates through the market stalls, her movements swift and precise. The camera quickly pans to capture her pursuer, a young man with wind-torn shorts and a low, playful determination. Dressed in a royal blue windbreaker and black gym shorts, he propels his black and silver skates forward with a powerful thrust across the uneven ground. They race in a sensual explosion of color and movement: stalls overflowing with red tomatoes and green vegetables, awnings draped with vibrant textiles, and the diverse crowd of shoppers reduced to a soft, impressionistic blur. The bright midday sun illuminates the scene, filtering through the gaps between the stalls and casting fast-moving, speckled patterns of light and shadow on the ground, creating an atmosphere of energy, joy, vitality, and youthful freedom.", + "seed": 20157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_7288_seed_20157_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": -0.0637578349094838, + "vlm_logprob": -0.0637578349094838, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.0637578349094838, + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2755281329154968 + ], + "opposite_logprob_tokens": [ + -1.6505281925201416 + ], + "logprob_reward": -0.2755281329154968, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2755281329154968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19714125990867615, + 0.06739718466997147, + 0.12123619765043259, + 0.09122506529092789, + 0.13558804988861084, + 0.09067507088184357, + 0.07604873925447464, + 0.06202353537082672, + 0.1082083210349083, + 0.050456564873456955 + ], + "credit_map": [ + 0.02763836272060871, + 0.017634058371186256, + 0.03869342431426048, + 0.027294309809803963, + 0.012152479961514473, + 0.016517294570803642, + 0.02680961973965168, + 0.030401714146137238, + 0.008131598122417927, + 0.007545081432908773, + 0.010518672876060009, + 0.008750613778829575, + 0.007132410537451506, + 0.00958787091076374, + 0.00876630749553442, + 0.006964623928070068, + 0.014802539721131325, + 0.011275634169578552, + 0.015170044265687466, + 0.014141619205474854, + 0.0175761878490448, + 0.010275131091475487, + 0.025069599971175194, + 0.012925439514219761, + 0.008299365639686584, + 0.012281953357160091, + 0.00886444840580225, + 0.014387535862624645, + 0.01446550153195858, + 0.00613279128447175, + 0.013191099278628826, + 0.013602372258901596, + 0.04313478618860245, + 0.009691556915640831, + 0.03564641252160072, + 0.014407718554139137, + 0.005683738272637129, + 0.006174035370349884, + 0.009379735216498375, + 0.011470065452158451, + 0.0067861368879675865, + 0.010219870135188103, + 0.01764155738055706, + 0.022556576877832413, + 0.007118239067494869, + 0.006340817082673311, + 0.007622153963893652, + 0.0123897148296237, + 0.007366072852164507, + 0.009433533996343613, + 0.010905980132520199, + 0.01046410109847784, + 0.008980397135019302, + 0.010474021546542645, + 0.008675293996930122, + 0.009749338962137699, + 0.00414299825206399, + 0.004511352628469467, + 0.011252145282924175, + 0.010885772295296192, + 0.005729389376938343, + 0.007521935272961855, + 0.008407097309827805, + 0.009572846814990044, + 0.011151702143251896, + 0.007324585225433111, + 0.018367798998951912, + 0.01355661079287529, + 0.0071573094464838505, + 0.014399230480194092, + 0.0209058690816164, + 0.01534520834684372, + 0.0037033844273537397, + 0.0037194816395640373, + 0.0069654593244194984, + 0.0068232533521950245, + 0.005528686568140984, + 0.0053589195013046265, + 0.00702726561576128, + 0.01133010908961296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.2755281329154968 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061477117240429 + ], + "opposite_logprob_tokens": [ + -3.8956148624420166 + ], + "logprob_reward": -0.02061477117240429, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061477117240429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37594395875930786, + 0.11143890768289566, + 0.0936320498585701, + 0.08796609938144684, + 0.05810381472110748, + 0.031676776707172394, + 0.07495534420013428, + 0.05345766246318817, + 0.045329250395298004, + 0.06749609857797623 + ], + "credit_map": [ + 0.05389111489057541, + 0.02583705633878708, + 0.0813322439789772, + 0.03875569999217987, + 0.022626426070928574, + 0.03571585193276405, + 0.042054787278175354, + 0.07573075592517853, + 0.010337652638554573, + 0.008609060198068619, + 0.014276711270213127, + 0.015342113561928272, + 0.011624608188867569, + 0.011000105179846287, + 0.02580391988158226, + 0.014444742351770401, + 0.0114048533141613, + 0.009534855373203754, + 0.008796095848083496, + 0.013935530558228493, + 0.012458905577659607, + 0.007740560919046402, + 0.01826115883886814, + 0.011500086635351181, + 0.007752502337098122, + 0.01168797817081213, + 0.01284096296876669, + 0.011193596757948399, + 0.01558891125023365, + 0.006193192210048437, + 0.009504910558462143, + 0.013204044662415981, + 0.01713741570711136, + 0.0034901765175163746, + 0.014570537954568863, + 0.0055221254006028175, + 0.002809901488944888, + 0.0032845367677509785, + 0.004567709285765886, + 0.006721411366015673, + 0.002441093325614929, + 0.0032945596612989902, + 0.007072430104017258, + 0.007545880042016506, + 0.0026107809972018003, + 0.0022001429460942745, + 0.00289277289994061, + 0.0036191176623106003, + 0.005679911933839321, + 0.008010565303266048, + 0.009725327603518963, + 0.00952887162566185, + 0.011013811454176903, + 0.008387473411858082, + 0.013531197793781757, + 0.009078183211386204, + 0.003919331822544336, + 0.0056605529971420765, + 0.0123452665284276, + 0.008716458454728127, + 0.003123733913525939, + 0.006207550875842571, + 0.005000832490622997, + 0.008483936078846455, + 0.004839044995605946, + 0.002972500864416361, + 0.008769994601607323, + 0.006062416825443506, + 0.0023428513668477535, + 0.008526766672730446, + 0.00433873338624835, + 0.007476939354091883, + 0.008440010249614716, + 0.00617192080244422, + 0.01129861082881689, + 0.010959775187075138, + 0.004331772215664387, + 0.00965727586299181, + 0.008162708953022957, + 0.008474022150039673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02061477117240429 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024942022282630205 + ], + "opposite_logprob_tokens": [ + -6.0024943351745605 + ], + "logprob_reward": -0.0024942022282630205, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024942022282630205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19188450276851654, + 0.13660959899425507, + 0.1811874806880951, + 0.06660018861293793, + 0.06825277209281921, + 0.09490054100751877, + 0.11876747012138367, + 0.06358633935451508, + 0.03989974036812782, + 0.03831132873892784 + ], + "credit_map": [ + 0.02184663712978363, + 0.014207996428012848, + 0.029553648084402084, + 0.024306247010827065, + 0.01424066349864006, + 0.027128539979457855, + 0.026469402015209198, + 0.03413137421011925, + 0.01474062167108059, + 0.018458707258105278, + 0.015882613137364388, + 0.025104450061917305, + 0.012673701159656048, + 0.009738395921885967, + 0.0246152114123106, + 0.015395904891192913, + 0.022073354572057724, + 0.02950911782681942, + 0.01778402365744114, + 0.029200667515397072, + 0.02365114726126194, + 0.013358153402805328, + 0.027855897322297096, + 0.017755132168531418, + 0.005820158869028091, + 0.009945063851773739, + 0.009912384673953056, + 0.010756419040262699, + 0.008667798712849617, + 0.006028860807418823, + 0.008850415237247944, + 0.00661908695474267, + 0.019235558807849884, + 0.005201814696192741, + 0.012272313237190247, + 0.00657616974785924, + 0.004461432807147503, + 0.004325338639318943, + 0.006662122905254364, + 0.009518022648990154, + 0.009852278046309948, + 0.007867186330258846, + 0.022108137607574463, + 0.025099782273173332, + 0.006058716215193272, + 0.0054247211664915085, + 0.005859385244548321, + 0.012630336917936802, + 0.0121105145663023, + 0.01789560727775097, + 0.01296289823949337, + 0.016317380592226982, + 0.015327244997024536, + 0.012222768738865852, + 0.016755590215325356, + 0.015175475738942623, + 0.004810264799743891, + 0.008006894960999489, + 0.011742917820811272, + 0.010396797209978104, + 0.006052026990801096, + 0.006689521484076977, + 0.007177478633821011, + 0.008710440248250961, + 0.006147576496005058, + 0.0029766452498733997, + 0.005730903707444668, + 0.004574133083224297, + 0.0041535841301083565, + 0.006626463029533625, + 0.0050112116150557995, + 0.004679222125560045, + 0.003963617607951164, + 0.0034027807414531708, + 0.005911433603614569, + 0.00422016903758049, + 0.003161374479532242, + 0.007747417315840721, + 0.004024777561426163, + 0.005879755597561598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0024942022282630205 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004089683759957552 + ], + "opposite_logprob_tokens": [ + -5.504089832305908 + ], + "logprob_reward": -0.004089683759957552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004089683759957552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15271949768066406, + 0.1879710853099823, + 0.20024019479751587, + 0.06959166377782822, + 0.0430573895573616, + 0.04681526497006416, + 0.1526462435722351, + 0.06567683815956116, + 0.037969816476106644, + 0.04331200197339058 + ], + "credit_map": [ + 0.02223271131515503, + 0.012011208571493626, + 0.030346767976880074, + 0.017693892121315002, + 0.009401834569871426, + 0.01166634913533926, + 0.020916808396577835, + 0.02844991348683834, + 0.017149923369288445, + 0.020784281194210052, + 0.022673221305012703, + 0.03469599038362503, + 0.016355548053979874, + 0.013912876136600971, + 0.04036456346511841, + 0.022034671157598495, + 0.026015393435955048, + 0.01952020265161991, + 0.018110115081071854, + 0.024636732414364815, + 0.026848385110497475, + 0.025853503495454788, + 0.033331990242004395, + 0.02592387981712818, + 0.0059327855706214905, + 0.013601639308035374, + 0.005644495598971844, + 0.012569226324558258, + 0.007705986965447664, + 0.006112608592957258, + 0.008372153155505657, + 0.00965277198702097, + 0.010466502048075199, + 0.0030942156445235014, + 0.005745055619627237, + 0.0046736751683056355, + 0.0030127798672765493, + 0.0036854480858892202, + 0.004887648392468691, + 0.007492064964026213, + 0.006178481038659811, + 0.0046286010183393955, + 0.006566147319972515, + 0.007054825779050589, + 0.00776823703199625, + 0.0046556079760193825, + 0.004816101863980293, + 0.005147265736013651, + 0.013676760718226433, + 0.017762182280421257, + 0.016753345727920532, + 0.01926366426050663, + 0.023400690406560898, + 0.01572531647980213, + 0.028183789923787117, + 0.01788049191236496, + 0.004913203418254852, + 0.005000223871320486, + 0.012325678952038288, + 0.00847494788467884, + 0.009613278321921825, + 0.007377823814749718, + 0.008666894398629665, + 0.009304788894951344, + 0.005875567439943552, + 0.0023348648101091385, + 0.006585386581718922, + 0.004080806393176317, + 0.003035684349015355, + 0.0055091772228479385, + 0.0050269803032279015, + 0.005521348677575588, + 0.00456110667437315, + 0.0045517729595303535, + 0.0054543460719287395, + 0.005057465750724077, + 0.003338223323225975, + 0.004527305252850056, + 0.0071557145565748215, + 0.00866607017815113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.004089683759957552 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062384471297264 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062384471297264, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062384471297264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1801026314496994, + 0.07850391417741776, + 0.12979423999786377, + 0.05465492606163025, + 0.10773056745529175, + 0.11936359852552414, + 0.15031102299690247, + 0.06795165687799454, + 0.07838336378335953, + 0.03320406377315521 + ], + "credit_map": [ + 0.022561004385352135, + 0.018295494839549065, + 0.02116548828780651, + 0.022906281054019928, + 0.014220700599253178, + 0.021444860845804214, + 0.03187237307429314, + 0.027636436745524406, + 0.009686678647994995, + 0.00846906378865242, + 0.016159147024154663, + 0.007101015653461218, + 0.006924032233655453, + 0.007361085619777441, + 0.013543488457798958, + 0.009259398095309734, + 0.020000120624899864, + 0.013899493031203747, + 0.013116809539496899, + 0.02013111114501953, + 0.019457677379250526, + 0.011309108696877956, + 0.017945457249879837, + 0.013934459537267685, + 0.004771838895976543, + 0.006611217744648457, + 0.009645527228713036, + 0.00927222240716219, + 0.006132296286523342, + 0.004454728681594133, + 0.005816070828586817, + 0.007951022125780582, + 0.02888176031410694, + 0.00837815273553133, + 0.023783493787050247, + 0.00992597360163927, + 0.007007478270679712, + 0.008723311126232147, + 0.00961003452539444, + 0.011420361697673798, + 0.01677931472659111, + 0.014824415557086468, + 0.017230018973350525, + 0.01974010467529297, + 0.01099518034607172, + 0.009230997413396835, + 0.01243860274553299, + 0.018124964088201523, + 0.01037929393351078, + 0.01841360703110695, + 0.018043523654341698, + 0.02213251031935215, + 0.012259138748049736, + 0.015161964111030102, + 0.023741265758872032, + 0.03017972782254219, + 0.005957083310931921, + 0.005792382638901472, + 0.013820909895002842, + 0.010263005271553993, + 0.00443527614697814, + 0.009343661367893219, + 0.007014689035713673, + 0.011324649676680565, + 0.009967876598238945, + 0.005268266890197992, + 0.013023996725678444, + 0.008927716873586178, + 0.005189879331737757, + 0.011518002487719059, + 0.014315599575638771, + 0.010172027163207531, + 0.00208836211822927, + 0.002561544068157673, + 0.00468550156801939, + 0.005263711325824261, + 0.0026497114449739456, + 0.005419460590928793, + 0.005061857867985964, + 0.005473912227898836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.016062384471297264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, clinging just inches above the worn cobblestones, fluidly weaves between two figures engaged in a high-speed roller-skating chase across a vibrant, crowded open-air market. The camera begins by following a young woman, her long brown hair tied in a high ponytail, flowing in the air as she glances back with a smile of excitement on her face. Dressed in a bright yellow tank top and denim shorts, she expertly navigates her scuffed burgundy roller skates through the market stalls, her movements swift and precise. The camera quickly pans to capture her pursuer, a young man with wind-torn shorts and a low, playful determination. Dressed in a royal blue windbreaker and black gym shorts, he propels his black and silver skates forward with a powerful thrust across the uneven ground. They race in a sensual explosion of color and movement: stalls overflowing with red tomatoes and green vegetables, awnings draped with vibrant textiles, and the diverse crowd of shoppers reduced to a soft, impressionistic blur. The bright midday sun illuminates the scene, filtering through the gaps between the stalls and casting fast-moving, speckled patterns of light and shadow on the ground, creating an atmosphere of energy, joy, vitality, and youthful freedom.", + "seed": 20158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_0/index_7288_seed_20158_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": -0.12000290742143989, + "vlm_logprob": -0.12000290742143989, + "vlm_yes_no_margin": 3.575000011920929, + "vlm_token_credit": -0.12000290742143989, + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5307173132896423 + ], + "opposite_logprob_tokens": [ + -1.030717372894287 + ], + "logprob_reward": -0.5307173132896423, + "yes_no_margin_reward": 0.5000000596046448, + "token_credit_reward": -0.5307173132896423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1754942685365677, + 0.06175592914223671, + 0.20806634426116943, + 0.08390052616596222, + 0.05696392059326172, + 0.06675301492214203, + 0.10550455749034882, + 0.08010367304086685, + 0.115435391664505, + 0.046022310853004456 + ], + "credit_map": [ + 0.023471225053071976, + 0.02810390666127205, + 0.029831325635313988, + 0.028719386085867882, + 0.010357696563005447, + 0.01483570970594883, + 0.017993085086345673, + 0.022181933745741844, + 0.009298319928348064, + 0.005314648151397705, + 0.005478400271385908, + 0.005581499543040991, + 0.003199083497747779, + 0.004675114061683416, + 0.02044285088777542, + 0.007766014896333218, + 0.02324107475578785, + 0.0258293729275465, + 0.0248695220798254, + 0.027234509587287903, + 0.02265336364507675, + 0.02952377125620842, + 0.03169027715921402, + 0.023024464026093483, + 0.005267367698252201, + 0.006240823771804571, + 0.013553495518863201, + 0.008695337921380997, + 0.009663739241659641, + 0.007266816217452288, + 0.017063476145267487, + 0.016149472445249557, + 0.00840743724256754, + 0.004823645576834679, + 0.0061218286864459515, + 0.008878201246261597, + 0.00677300151437521, + 0.00526460213586688, + 0.008360733278095722, + 0.008334469050168991, + 0.005708703305572271, + 0.008042952045798302, + 0.014429658651351929, + 0.012392260134220123, + 0.006515280343592167, + 0.005253957584500313, + 0.006761705968528986, + 0.0076484982855618, + 0.007792247459292412, + 0.01303123589605093, + 0.01352945901453495, + 0.015856381505727768, + 0.013689874671399593, + 0.016164615750312805, + 0.011913234367966652, + 0.013527496717870235, + 0.009338168427348137, + 0.006726656574755907, + 0.012202880345284939, + 0.010937058366835117, + 0.010726185515522957, + 0.009395788423717022, + 0.010400436818599701, + 0.01037649903446436, + 0.008377337828278542, + 0.009110361337661743, + 0.011004423722624779, + 0.027943499386310577, + 0.0073257023468613625, + 0.01110836025327444, + 0.027869585901498795, + 0.012696116231381893, + 0.0037744129076600075, + 0.007270596921443939, + 0.004188974853605032, + 0.007619801443070173, + 0.003359843511134386, + 0.004767767619341612, + 0.00933906901627779, + 0.005701844580471516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.5307173132896423 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374604508280754 + ], + "opposite_logprob_tokens": [ + -3.4087460041046143 + ], + "logprob_reward": -0.03374604508280754, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374604508280754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11212042719125748, + 0.17419126629829407, + 0.14579087495803833, + 0.1288238763809204, + 0.03698582202196121, + 0.1296149492263794, + 0.0737847313284874, + 0.06467795372009277, + 0.08151008933782578, + 0.052500031888484955 + ], + "credit_map": [ + 0.014255463145673275, + 0.017596988007426262, + 0.012888127006590366, + 0.011866151355206966, + 0.013214779086411, + 0.015016901306807995, + 0.014337417669594288, + 0.012944594956934452, + 0.02780015580356121, + 0.011042420752346516, + 0.016262255609035492, + 0.014174743555486202, + 0.009501222521066666, + 0.015078574419021606, + 0.05779460817575455, + 0.022537291049957275, + 0.014152566902339458, + 0.01733234152197838, + 0.018407881259918213, + 0.019828559830784798, + 0.020009230822324753, + 0.016108646988868713, + 0.023244649171829224, + 0.016707008704543114, + 0.010264357551932335, + 0.013004905544221401, + 0.02158064767718315, + 0.012550394982099533, + 0.014266857877373695, + 0.01233108714222908, + 0.025170255452394485, + 0.019655359908938408, + 0.0047524976544082165, + 0.0031110839918255806, + 0.0038732520770281553, + 0.004242078401148319, + 0.005565023515373468, + 0.003497019177302718, + 0.006628892850130796, + 0.005315972492098808, + 0.008593400940299034, + 0.013213682919740677, + 0.024274662137031555, + 0.018266044557094574, + 0.013433338142931461, + 0.02282809466123581, + 0.01351134292781353, + 0.015494368970394135, + 0.006202139891684055, + 0.0077479081228375435, + 0.01258760318160057, + 0.01505929697304964, + 0.011552262119948864, + 0.0043579894118011, + 0.011849296279251575, + 0.004428233951330185, + 0.010278578847646713, + 0.011625297367572784, + 0.00970229972153902, + 0.010748621076345444, + 0.00463969586417079, + 0.004474746994674206, + 0.00814315490424633, + 0.005065554287284613, + 0.0063629974611103535, + 0.006737221032381058, + 0.009925018064677715, + 0.01657329499721527, + 0.0084678390994668, + 0.006375722587108612, + 0.017925867810845375, + 0.009142126888036728, + 0.004309211391955614, + 0.009955712594091892, + 0.0051954360678792, + 0.007533642463386059, + 0.004268669988960028, + 0.0065767718479037285, + 0.009495850652456284, + 0.005164733622223139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.03374604508280754 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648118279874325 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004648118279874325, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648118279874325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12720023095607758, + 0.0794156864285469, + 0.0597999282181263, + 0.08576397597789764, + 0.08771345764398575, + 0.03465103358030319, + 0.3924455940723419, + 0.07382782548666, + 0.028394188731908798, + 0.03078807145357132 + ], + "credit_map": [ + 0.0156998373568058, + 0.022440800443291664, + 0.020836446434259415, + 0.023172803223133087, + 0.01004940364509821, + 0.009409938007593155, + 0.011075974442064762, + 0.014515035785734653, + 0.012291372753679752, + 0.0055137206800282, + 0.007114908192306757, + 0.009124157018959522, + 0.004209347069263458, + 0.006289780139923096, + 0.024818072095513344, + 0.010054332204163074, + 0.006429854314774275, + 0.005286168772727251, + 0.007554375566542149, + 0.007725514937192202, + 0.00684934388846159, + 0.007290094159543514, + 0.01379187311977148, + 0.004872703459113836, + 0.00825962983071804, + 0.009031828492879868, + 0.023368557915091515, + 0.008501624688506126, + 0.011229259893298149, + 0.0071662357077002525, + 0.010274963453412056, + 0.007931875996291637, + 0.0117914704605937, + 0.008643794804811478, + 0.007922329008579254, + 0.0092404680326581, + 0.018786655738949776, + 0.009194598533213139, + 0.01268982607871294, + 0.009444317780435085, + 0.002993449568748474, + 0.004695878829807043, + 0.004731974098831415, + 0.0075539532117545605, + 0.0023633898235857487, + 0.002930002985522151, + 0.003980521112680435, + 0.0054018693044781685, + 0.026574915274977684, + 0.057288166135549545, + 0.06306261569261551, + 0.06587708741426468, + 0.04452785477042198, + 0.0331357941031456, + 0.06408871710300446, + 0.0378904715180397, + 0.005061001982539892, + 0.006252847611904144, + 0.009072836488485336, + 0.01074003241956234, + 0.006129480432718992, + 0.008996916003525257, + 0.018517840653657913, + 0.009056874550879002, + 0.002428679261356592, + 0.0029536248184740543, + 0.0033766215201467276, + 0.004273377824574709, + 0.0037261706311255693, + 0.0033186764921993017, + 0.0054677813313901424, + 0.0028492591809481382, + 0.0026759752072393894, + 0.0036971140652894974, + 0.004560676869004965, + 0.003915466368198395, + 0.0020939006935805082, + 0.003177024656906724, + 0.006667064502835274, + 0.004000851884484291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.004648118279874325 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622324395924807 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.007622324395924807, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622324395924807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13185086846351624, + 0.0504712238907814, + 0.03238410875201225, + 0.11464115977287292, + 0.16081471741199493, + 0.03479638695716858, + 0.3798827528953552, + 0.03486165031790733, + 0.032366037368774414, + 0.027931122109293938 + ], + "credit_map": [ + 0.01843550242483616, + 0.014366477727890015, + 0.020108351483941078, + 0.026019608601927757, + 0.01334421057254076, + 0.010715709999203682, + 0.017048846930265427, + 0.011812154203653336, + 0.006352412514388561, + 0.00448259012773633, + 0.005558353383094072, + 0.006311577744781971, + 0.0037754285149276257, + 0.005416046362370253, + 0.01232596579939127, + 0.0062488471157848835, + 0.0036941661965101957, + 0.004161699675023556, + 0.004391571041196585, + 0.004370799288153648, + 0.002833291422575712, + 0.004204300232231617, + 0.004509173799306154, + 0.004219104070216417, + 0.007832422852516174, + 0.011179354973137379, + 0.020662210881710052, + 0.01237276941537857, + 0.015115258283913136, + 0.011171102523803711, + 0.019217222929000854, + 0.017090817913413048, + 0.015137546695768833, + 0.012756739743053913, + 0.010647560469806194, + 0.014636551029980183, + 0.03627233952283859, + 0.018500657752156258, + 0.0280019398778677, + 0.02486138418316841, + 0.0026153528597205877, + 0.0047642504796385765, + 0.005733151454478502, + 0.00528633501380682, + 0.00285159470513463, + 0.0044181435368955135, + 0.0036800189409404993, + 0.005447541829198599, + 0.025095120072364807, + 0.039745744317770004, + 0.06526148319244385, + 0.05589437484741211, + 0.04042665660381317, + 0.04686921834945679, + 0.0639251247048378, + 0.0426650308072567, + 0.0033595517743378878, + 0.0028268140740692616, + 0.004780053626745939, + 0.004594328347593546, + 0.0037295448128134012, + 0.0035604953300207853, + 0.008220909163355827, + 0.0037899527233093977, + 0.0027893371880054474, + 0.00355387176387012, + 0.0033230693079531193, + 0.0055569615215063095, + 0.0035786570515483618, + 0.004492262378334999, + 0.005181378684937954, + 0.0038905006367713213, + 0.0024760053493082523, + 0.003542770631611347, + 0.0026363220531493425, + 0.003587427781894803, + 0.002321898238733411, + 0.004896005615592003, + 0.0051944381557404995, + 0.00327625242061913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.007622324395924807 + }, + { + "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, just inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vibrant, crowded open-air market with stalls overflowing with red tomatoes and green vegetables?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young woman visible wearing a bright yellow tank top, denim shorts, and scuffed burgundy roller skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young man visible wearing a royal blue windbreaker, black gym shorts, and black and silver skates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera fluidly weave between the two figures while tracking their high-speed roller-skating chase?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24322932958602905, + 0.07040570676326752, + 0.06147358939051628, + 0.118545301258564, + 0.08387621492147446, + 0.07307466864585876, + 0.18539518117904663, + 0.0715879276394844, + 0.050447016954422, + 0.04196513444185257 + ], + "credit_map": [ + 0.04160813242197037, + 0.03459274768829346, + 0.0594002902507782, + 0.027725204825401306, + 0.013582549057900906, + 0.013777331449091434, + 0.025253625586628914, + 0.02728942036628723, + 0.008740492165088654, + 0.009883876889944077, + 0.011969509534537792, + 0.007935873232781887, + 0.007108933757990599, + 0.0055755735374987125, + 0.009825960732996464, + 0.009365473873913288, + 0.006362576503306627, + 0.006265138741582632, + 0.007637070957571268, + 0.006800472270697355, + 0.006237538531422615, + 0.012025467120110989, + 0.009645900689065456, + 0.006499418988823891, + 0.007539127487689257, + 0.010415484197437763, + 0.014740643091499805, + 0.01144714467227459, + 0.013366052880883217, + 0.011219820007681847, + 0.026616260409355164, + 0.023200757801532745, + 0.010616595856845379, + 0.0049911304377019405, + 0.007838491350412369, + 0.005905786529183388, + 0.008614377118647099, + 0.006503244861960411, + 0.019215784966945648, + 0.02019079402089119, + 0.0070559438318014145, + 0.010535637848079205, + 0.009963784366846085, + 0.01205164659768343, + 0.008075643330812454, + 0.006265651434659958, + 0.010245708748698235, + 0.008880640380084515, + 0.015267033129930496, + 0.024316910654306412, + 0.02915848419070244, + 0.037191521376371384, + 0.02367527410387993, + 0.015941567718982697, + 0.02535839192569256, + 0.014485983178019524, + 0.005329914391040802, + 0.004833200015127659, + 0.008050410076975822, + 0.009579729288816452, + 0.0077313026413321495, + 0.007987025193870068, + 0.014717628248035908, + 0.013358712196350098, + 0.0047338129952549934, + 0.00462370878085494, + 0.004943800158798695, + 0.008220583200454712, + 0.004809249192476273, + 0.007883956655859947, + 0.008751981891691685, + 0.006479921285063028, + 0.004381157923489809, + 0.006078153382986784, + 0.004776848945766687, + 0.006932797841727734, + 0.0031678639352321625, + 0.00451772240921855, + 0.0070344009436666965, + 0.005076186265796423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.023280736058950424 + } + ], + "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/0000060/group_1/index_tensor([9275, 9275])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_tensor([9275, 9275])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb4240bc1af4fea6af943776eda5359e6a737b72 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_tensor([9275, 9275])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Amidst a heavy nighttime snowfall, a high-angle, wide-angle shot slowly sweeps left across a vast, floodlit football stadium. The camera captures the entire spectacle, showing a field almost entirely covered in pristine white snow, with barely a hint of green turf beneath. Thick, swirling snowflakes drift relentlessly down, caught by the powerful, cold white light from the overhead stadium lights, casting long, dramatic shadows across the field. The tribune is packed with a dense, roaring crowd, crammed with spectators clad in heavy winter coats, colorful woolen hats, and team scarves. Their collective breath hangs like a visible mist in the chilly air. On the field below, players, clad in vibrant, solid jerseys—one team all-red, the other all-blue—move with energy and determination, vying for a bright orange ball. They sprint, slide, and tackle, kicking up explosive plumes of snow with every powerful stride and challenge, creating a scene of intense, raw athletic passion that contrasts with the harsh winter weather and creates a thrilling atmosphere.", + "seed": 6270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6270_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": -0.055766853410750626, + "vlm_logprob": -0.055766853410750626, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.055766853410750626, + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04168618470430374 + ], + "opposite_logprob_tokens": [ + -3.2916862964630127 + ], + "logprob_reward": -0.04168618470430374, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04168618470430374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5019312500953674, + 0.06472094357013702, + 0.03993825986981392, + 0.0479482002556324, + 0.04251309484243393, + 0.027112344279885292, + 0.03442039713263512, + 0.04738368093967438, + 0.17530125379562378, + 0.01873047649860382 + ], + "credit_map": [ + 0.04366850480437279, + 0.05855775251984596, + 0.08886464685201645, + 0.07588688284158707, + 0.04691087454557419, + 0.07763241231441498, + 0.050468627363443375, + 0.05994165316224098, + 0.007273724302649498, + 0.006720447447150946, + 0.012073089368641376, + 0.007348434068262577, + 0.009089609608054161, + 0.008508686907589436, + 0.006763771176338196, + 0.006943186745047569, + 0.0037327941972762346, + 0.004142796620726585, + 0.004847948905080557, + 0.004487700294703245, + 0.006209163926541805, + 0.0053651584312319756, + 0.006319576874375343, + 0.004833124577999115, + 0.004891813267022371, + 0.006094394717365503, + 0.005794812925159931, + 0.004878953564912081, + 0.00937735941261053, + 0.005338860210031271, + 0.006124959792941809, + 0.0054470538161695, + 0.003034858498722315, + 0.008720744401216507, + 0.0047800843603909016, + 0.004571465775370598, + 0.007880687713623047, + 0.0052349865436553955, + 0.005003969185054302, + 0.003286303486675024, + 0.0017878302605822682, + 0.003953012637794018, + 0.004602293949574232, + 0.003629527986049652, + 0.003997630439698696, + 0.0034437496215105057, + 0.0031960969790816307, + 0.002502203918993473, + 0.003307138103991747, + 0.004339582286775112, + 0.0045325178653001785, + 0.004566522315144539, + 0.005769169889390469, + 0.0044539389200508595, + 0.003769905539229512, + 0.0036816245410591364, + 0.0028223793487995863, + 0.012612390331923962, + 0.0033631923142820597, + 0.004386467393487692, + 0.01124478317797184, + 0.0043721916154026985, + 0.004516391083598137, + 0.004065889399498701, + 0.00421536061912775, + 0.07518227398395538, + 0.0022433351259678602, + 0.004815146792680025, + 0.07617748528718948, + 0.0064603709615767, + 0.0029625934548676014, + 0.0032446777913719416, + 0.0020552990026772022, + 0.0026378801558166742, + 0.0018321264069527388, + 0.0029741530306637287, + 0.0022899939212948084, + 0.0018142372136935592, + 0.002502608811482787, + 0.0026241799350827932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04168618470430374 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030654817819595 + ], + "opposite_logprob_tokens": [ + -2.350306510925293 + ], + "logprob_reward": -0.10030654817819595, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030654817819595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1652323603630066, + 0.07016338407993317, + 0.08335793018341064, + 0.07916777580976486, + 0.14487868547439575, + 0.021552588790655136, + 0.054737579077482224, + 0.030004657804965973, + 0.3064577579498291, + 0.044447269290685654 + ], + "credit_map": [ + 0.018753893673419952, + 0.01788156107068062, + 0.01813424378633499, + 0.01598377153277397, + 0.026154669001698494, + 0.03073607198894024, + 0.012151326052844524, + 0.02543681673705578, + 0.0063837068155407906, + 0.007454427890479565, + 0.009193469770252705, + 0.008884369395673275, + 0.011211163364350796, + 0.010638915002346039, + 0.008519969880580902, + 0.007877358235418797, + 0.007843720726668835, + 0.015300722792744637, + 0.008764763362705708, + 0.013877823017537594, + 0.0131070576608181, + 0.008543194271624088, + 0.010887944139540195, + 0.005032707471400499, + 0.005792887415736914, + 0.01893441192805767, + 0.009961847215890884, + 0.00766395078971982, + 0.017534049227833748, + 0.006713952403515577, + 0.008027279749512672, + 0.004539398476481438, + 0.009794509969651699, + 0.04620134457945824, + 0.009696362540125847, + 0.012843417935073376, + 0.041807882487773895, + 0.011152886785566807, + 0.008145580999553204, + 0.0052366903983056545, + 0.00120642373804003, + 0.003481812309473753, + 0.0028086970560252666, + 0.002751468913629651, + 0.004041227977722883, + 0.0028510908596217632, + 0.0028144761454313993, + 0.001597392139956355, + 0.004646040964871645, + 0.007775378413498402, + 0.007858200930058956, + 0.01014594454318285, + 0.008356940001249313, + 0.005555361043661833, + 0.005139584187418222, + 0.005260128993541002, + 0.0028788878116756678, + 0.005902417935431004, + 0.0029142708517611027, + 0.0034740206319838762, + 0.005662209820002317, + 0.002644514199346304, + 0.003400949062779546, + 0.00312738842330873, + 0.009868775494396687, + 0.11970660090446472, + 0.008158975280821323, + 0.019148891791701317, + 0.11922144144773483, + 0.011301807127892971, + 0.009541427716612816, + 0.0095098577439785, + 0.0033851880580186844, + 0.005870976485311985, + 0.0037710207980126143, + 0.006394447758793831, + 0.006512187886983156, + 0.003173164324834943, + 0.007333522662520409, + 0.008006757125258446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.10030654817819595 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.3882831931114197, + 0.11304047703742981, + 0.09814188629388809, + 0.03869118168950081, + 0.13454760611057281, + 0.03222920745611191, + 0.028095681220293045, + 0.0388929508626461, + 0.10423408448696136, + 0.02384374476969242 + ], + "credit_map": [ + 0.03438888490200043, + 0.04232336953282356, + 0.06756255775690079, + 0.07764063775539398, + 0.03755864128470421, + 0.04416796565055847, + 0.039193443953990936, + 0.0454476997256279, + 0.00866667926311493, + 0.011448971927165985, + 0.025108031928539276, + 0.012486064806580544, + 0.017194224521517754, + 0.01398493628948927, + 0.012475172057747841, + 0.011676406487822533, + 0.008932694792747498, + 0.012004596181213856, + 0.015442654490470886, + 0.016904620453715324, + 0.014052030630409718, + 0.010714816860854626, + 0.011754612438380718, + 0.00833586324006319, + 0.0031794558744877577, + 0.005520512815564871, + 0.004535442218184471, + 0.006277690175920725, + 0.005428076256066561, + 0.005336164962500334, + 0.004746058955788612, + 0.0036677848547697067, + 0.009254447184503078, + 0.03691573441028595, + 0.014258351176977158, + 0.014859683811664581, + 0.03370657563209534, + 0.011327847838401794, + 0.007223410531878471, + 0.007001576479524374, + 0.002878406783565879, + 0.003650728380307555, + 0.004464686848223209, + 0.004045209381729364, + 0.004861928056925535, + 0.005327856168150902, + 0.004506655037403107, + 0.0024937421549111605, + 0.001796620897948742, + 0.0036187940277159214, + 0.0034845820628106594, + 0.00559519836679101, + 0.003969117999076843, + 0.0028449841775000095, + 0.0041479491628706455, + 0.002638437319546938, + 0.003859605873003602, + 0.007621650118380785, + 0.0034197773784399033, + 0.005335897672921419, + 0.006589998956769705, + 0.005355524830520153, + 0.003720474662259221, + 0.0029900248628109694, + 0.004838948138058186, + 0.03479316085577011, + 0.00458949152380228, + 0.010609475895762444, + 0.03567104786634445, + 0.005247689317911863, + 0.004489092156291008, + 0.003995186183601618, + 0.0023515482898801565, + 0.0020185192115604877, + 0.0033079476561397314, + 0.004252607934176922, + 0.002946762368083, + 0.001733324839733541, + 0.0030644468497484922, + 0.00416859146207571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.006738320458680391 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032132267951965 + ], + "opposite_logprob_tokens": [ + -2.3503212928771973 + ], + "logprob_reward": -0.10032132267951965, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032132267951965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32084283232688904, + 0.12294560670852661, + 0.04280281066894531, + 0.05981150269508362, + 0.04860041290521622, + 0.04391898587346077, + 0.03329963609576225, + 0.04474594444036484, + 0.22916743159294128, + 0.05386493355035782 + ], + "credit_map": [ + 0.03045227937400341, + 0.03306935727596283, + 0.048532117158174515, + 0.05472791939973831, + 0.03932557255029678, + 0.03936190530657768, + 0.038734883069992065, + 0.03663874790072441, + 0.01132927555590868, + 0.013111714273691177, + 0.021878035739064217, + 0.014577923342585564, + 0.016077719628810883, + 0.013769138604402542, + 0.021244047209620476, + 0.010957730934023857, + 0.005834212061017752, + 0.004703807178884745, + 0.005941089708358049, + 0.003618402173742652, + 0.008315145038068295, + 0.005174040794372559, + 0.004644404631108046, + 0.004571704659610987, + 0.005890915170311928, + 0.00785907357931137, + 0.009542829357087612, + 0.00799703598022461, + 0.009553101845085621, + 0.006113960407674313, + 0.007569620385766029, + 0.005284957587718964, + 0.00383947417140007, + 0.006158835254609585, + 0.010561992414295673, + 0.006725119426846504, + 0.006441203877329826, + 0.0053232815116643906, + 0.005870872642844915, + 0.003679628949612379, + 0.002699323697015643, + 0.010245717130601406, + 0.00463740061968565, + 0.0033847996965050697, + 0.010692963376641273, + 0.005491015035659075, + 0.004002366214990616, + 0.00276539521291852, + 0.0023865175899118185, + 0.003957604058086872, + 0.006001625210046768, + 0.006475234404206276, + 0.003981958609074354, + 0.0035688895732164383, + 0.0032576450612396, + 0.003670157864689827, + 0.004921902436763048, + 0.006012313533574343, + 0.004262974951416254, + 0.0071158274076879025, + 0.006413818337023258, + 0.006332744378596544, + 0.006752489134669304, + 0.0029338656459003687, + 0.006735406816005707, + 0.0911126434803009, + 0.006442997604608536, + 0.011661490425467491, + 0.09079395979642868, + 0.0087417121976614, + 0.006276325322687626, + 0.0074028633534908295, + 0.005485231522470713, + 0.007309569511562586, + 0.004367514047771692, + 0.007836038246750832, + 0.006695886142551899, + 0.005506196990609169, + 0.006170499138534069, + 0.010493991896510124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.10032132267951965 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781891033053398 + ], + "opposite_logprob_tokens": [ + -3.5297818183898926 + ], + "logprob_reward": -0.029781891033053398, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781891033053398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35909122228622437, + 0.10001049190759659, + 0.06353671103715897, + 0.0468834824860096, + 0.15202519297599792, + 0.06311818957328796, + 0.023038988932967186, + 0.028109045699238777, + 0.14702950417995453, + 0.017157165333628654 + ], + "credit_map": [ + 0.028430255129933357, + 0.04223223775625229, + 0.04684155434370041, + 0.056853506714105606, + 0.038726046681404114, + 0.04966685548424721, + 0.038086842745542526, + 0.05825392156839371, + 0.009001056663691998, + 0.015018485486507416, + 0.015093766152858734, + 0.007378794252872467, + 0.013944242149591446, + 0.01185031421482563, + 0.019172757863998413, + 0.008551076985895634, + 0.005263519939035177, + 0.010010373778641224, + 0.007553548086434603, + 0.005980508867651224, + 0.00956619530916214, + 0.008934770710766315, + 0.010476271621882915, + 0.005751525983214378, + 0.0033849950414150953, + 0.0048599704168736935, + 0.007067390251904726, + 0.011063032783567905, + 0.00473956111818552, + 0.0049293506890535355, + 0.00691163819283247, + 0.003927541896700859, + 0.009545407257974148, + 0.045308541506528854, + 0.013123825192451477, + 0.0172781553119421, + 0.04034610465168953, + 0.011669445782899857, + 0.008388022892177105, + 0.006365695968270302, + 0.0045400806702673435, + 0.01467984076589346, + 0.006476063746958971, + 0.006670002359896898, + 0.012993186712265015, + 0.005401344504207373, + 0.007947339676320553, + 0.004410331137478352, + 0.002159727271646261, + 0.003004810307174921, + 0.003187598427757621, + 0.0033620265312492847, + 0.002885716035962105, + 0.003078937530517578, + 0.002586070680990815, + 0.002774100983515382, + 0.0028517963364720345, + 0.0032103825360536575, + 0.003017741022631526, + 0.0041272626258432865, + 0.003490220056846738, + 0.004006735514849424, + 0.004207764286547899, + 0.003197143552824855, + 0.005201461724936962, + 0.056483566761016846, + 0.005672167055308819, + 0.00811153557151556, + 0.05591730773448944, + 0.0058867717161774635, + 0.004528052173554897, + 0.0052286479622125626, + 0.002003495115786791, + 0.0016386302886530757, + 0.0018576870206743479, + 0.0028302182909101248, + 0.0023778167087584734, + 0.0013988264836370945, + 0.0020425112452358007, + 0.003007981227710843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.029781891033053398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a heavy nighttime snowfall, a high-angle, wide-angle shot slowly sweeps left across a vast, floodlit football stadium. The camera captures the entire spectacle, showing a field almost entirely covered in pristine white snow, with barely a hint of green turf beneath. Thick, swirling snowflakes drift relentlessly down, caught by the powerful, cold white light from the overhead stadium lights, casting long, dramatic shadows across the field. The tribune is packed with a dense, roaring crowd, crammed with spectators clad in heavy winter coats, colorful woolen hats, and team scarves. Their collective breath hangs like a visible mist in the chilly air. On the field below, players, clad in vibrant, solid jerseys—one team all-red, the other all-blue—move with energy and determination, vying for a bright orange ball. They sprint, slide, and tackle, kicking up explosive plumes of snow with every powerful stride and challenge, creating a scene of intense, raw athletic passion that contrasts with the harsh winter weather and creates a thrilling atmosphere.", + "seed": 6271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6271_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": -0.038549369759857656, + "vlm_logprob": -0.038549369759857656, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.038549369759857656, + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022767921909689903 + ], + "opposite_logprob_tokens": [ + -3.897768020629883 + ], + "logprob_reward": -0.022767921909689903, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022767921909689903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2246183454990387, + 0.025974268093705177, + 0.07881515473127365, + 0.06368570774793625, + 0.05023973807692528, + 0.08853844553232193, + 0.1913072019815445, + 0.16799819469451904, + 0.06896310299634933, + 0.03985985741019249 + ], + "credit_map": [ + 0.021407511085271835, + 0.02020207606256008, + 0.04489120468497276, + 0.03340630233287811, + 0.037607111036777496, + 0.015684114769101143, + 0.03249421715736389, + 0.018925804644823074, + 0.0025626234710216522, + 0.002511163940653205, + 0.005521190352737904, + 0.003902614116668701, + 0.003110966645181179, + 0.001744302804581821, + 0.003906317055225372, + 0.002715090988203883, + 0.007864058949053288, + 0.00845393631607294, + 0.008381245657801628, + 0.012337569147348404, + 0.01039196364581585, + 0.007908441126346588, + 0.01143199298530817, + 0.01204594410955906, + 0.005388849414885044, + 0.008846422657370567, + 0.01460836362093687, + 0.005866391584277153, + 0.007013499271124601, + 0.00749099999666214, + 0.00644228421151638, + 0.008028893731534481, + 0.0043512252159416676, + 0.004539042711257935, + 0.005977295804768801, + 0.008092460222542286, + 0.0055977459996938705, + 0.004971022251993418, + 0.005835029296576977, + 0.010875914245843887, + 0.007635120768100023, + 0.007568588480353355, + 0.0098088588565588, + 0.008813965134322643, + 0.023398755118250847, + 0.009050289168953896, + 0.0059449379332363605, + 0.016317931935191154, + 0.016525622457265854, + 0.023938464000821114, + 0.009759931825101376, + 0.007682918105274439, + 0.06647122651338577, + 0.02066095732152462, + 0.014749650843441486, + 0.031518418341875076, + 0.013092389330267906, + 0.01844598539173603, + 0.011861619539558887, + 0.013283127918839455, + 0.02560066059231758, + 0.050316788256168365, + 0.01923007145524025, + 0.016167540103197098, + 0.005559277720749378, + 0.011725903488695621, + 0.004875118378549814, + 0.009733580984175205, + 0.009090681560337543, + 0.008216413669288158, + 0.010817701928317547, + 0.00894442293792963, + 0.004675367847084999, + 0.0035501664970070124, + 0.008790471591055393, + 0.004053972661495209, + 0.0026362016797065735, + 0.00439344672486186, + 0.006970334332436323, + 0.004789894446730614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.022767921909689903 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10961701720952988, + 0.11162687093019485, + 0.09772650897502899, + 0.11443181335926056, + 0.06344880908727646, + 0.05617167055606842, + 0.13320529460906982, + 0.10080048441886902, + 0.07091355323791504, + 0.14205798506736755 + ], + "credit_map": [ + 0.011056825518608093, + 0.012509243562817574, + 0.015171680599451065, + 0.01193265151232481, + 0.014610902406275272, + 0.0076356204226613045, + 0.020776944234967232, + 0.01592315174639225, + 0.009726008400321007, + 0.011429721489548683, + 0.011784485541284084, + 0.016779279336333275, + 0.012745874002575874, + 0.008145664818584919, + 0.024499669671058655, + 0.01651616580784321, + 0.00810899119824171, + 0.012477701529860497, + 0.0085968729108572, + 0.01372072845697403, + 0.008437355048954487, + 0.011065659113228321, + 0.012897380627691746, + 0.02242182567715645, + 0.009189686737954617, + 0.018068820238113403, + 0.01889970898628235, + 0.012350101955235004, + 0.011056714691221714, + 0.015924174338579178, + 0.01301487348973751, + 0.015927733853459358, + 0.005580139812082052, + 0.00420953007414937, + 0.008849047124385834, + 0.012984286062419415, + 0.005467821843922138, + 0.006353345699608326, + 0.010227003134787083, + 0.009777638129889965, + 0.005757004022598267, + 0.005012127570807934, + 0.006440118420869112, + 0.006292335223406553, + 0.011858992278575897, + 0.007685641758143902, + 0.005449111573398113, + 0.007676340639591217, + 0.0109030082821846, + 0.022151509299874306, + 0.004397249314934015, + 0.004880786873400211, + 0.05075347051024437, + 0.014455665834248066, + 0.006769711151719093, + 0.018893886357545853, + 0.007169278338551521, + 0.01082667987793684, + 0.008715660311281681, + 0.011298296973109245, + 0.014568954706192017, + 0.020573362708091736, + 0.017772136256098747, + 0.009876114316284657, + 0.00553047563880682, + 0.006700366735458374, + 0.006434143055230379, + 0.017037032172083855, + 0.007304058410227299, + 0.007194281555712223, + 0.010123605839908123, + 0.010589594952762127, + 0.014907989650964737, + 0.012926137074828148, + 0.007840440608561039, + 0.006621488835662603, + 0.013959108851850033, + 0.046914104372262955, + 0.015308757312595844, + 0.023579958826303482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.038097646087408066 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3434796929359436, + 0.08147932589054108, + 0.05920769274234772, + 0.06511040031909943, + 0.034688208252191544, + 0.03415362164378166, + 0.0558190755546093, + 0.13194292783737183, + 0.08571789413690567, + 0.10840112715959549 + ], + "credit_map": [ + 0.02839730493724346, + 0.02997123822569847, + 0.06353414058685303, + 0.05600614845752716, + 0.04685450717806816, + 0.023355059325695038, + 0.058567535132169724, + 0.0367937795817852, + 0.0063925632275640965, + 0.008603065274655819, + 0.011249429546296597, + 0.0124168386682868, + 0.008015827275812626, + 0.005335592664778233, + 0.01700194552540779, + 0.012464062310755253, + 0.005234560929238796, + 0.006286316551268101, + 0.00764291500672698, + 0.008411796763539314, + 0.006656727287918329, + 0.005690311547368765, + 0.007966455072164536, + 0.011318609118461609, + 0.005168247502297163, + 0.009167199954390526, + 0.015435442328453064, + 0.00684695178642869, + 0.006335275713354349, + 0.006801873445510864, + 0.006764546502381563, + 0.00859086960554123, + 0.0038199443370103836, + 0.003939240239560604, + 0.0048916880041360855, + 0.004469591658562422, + 0.003902291413396597, + 0.00558473402634263, + 0.004190872889012098, + 0.003889841726049781, + 0.002837123814970255, + 0.0033765872940421104, + 0.002931229304522276, + 0.0033635790459811687, + 0.010160920210182667, + 0.003963617607951164, + 0.0026691369712352753, + 0.004851427860558033, + 0.004261409398168325, + 0.007074570748955011, + 0.0027988844085484743, + 0.004378384444862604, + 0.018068434670567513, + 0.005561572499573231, + 0.004927912726998329, + 0.008747907355427742, + 0.010449279099702835, + 0.014230308122932911, + 0.00867594312876463, + 0.00880430918186903, + 0.0188496895134449, + 0.042963944375514984, + 0.011340716853737831, + 0.01662873663008213, + 0.008468478918075562, + 0.006092802621424198, + 0.008834611624479294, + 0.016418635845184326, + 0.007070033811032772, + 0.011125904507935047, + 0.014940470457077026, + 0.012766949832439423, + 0.01227654330432415, + 0.009461220353841782, + 0.013013260439038277, + 0.007382966112345457, + 0.009582923725247383, + 0.030054451897740364, + 0.011625532060861588, + 0.015004239045083523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.00525094848126173 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030201822519302 + ], + "opposite_logprob_tokens": [ + -2.35030198097229 + ], + "logprob_reward": -0.10030201822519302, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030201822519302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3648107945919037, + 0.19099049270153046, + 0.0890936329960823, + 0.04679707810282707, + 0.022543109953403473, + 0.03576989099383354, + 0.0720091238617897, + 0.07356230914592743, + 0.043810002505779266, + 0.060613568872213364 + ], + "credit_map": [ + 0.029680052772164345, + 0.03904304653406143, + 0.045472051948308945, + 0.04990847036242485, + 0.05518798902630806, + 0.022527040913701057, + 0.08166251331567764, + 0.04132961109280586, + 0.015280349180102348, + 0.01784844510257244, + 0.017536252737045288, + 0.025837760418653488, + 0.020680252462625504, + 0.012526324950158596, + 0.0489494651556015, + 0.032331645488739014, + 0.007092075888067484, + 0.008927675895392895, + 0.008589453063905239, + 0.014424430206418037, + 0.01239542756229639, + 0.008880853652954102, + 0.01378636620938778, + 0.014997339807450771, + 0.0038175764493644238, + 0.006650844588875771, + 0.01070831436663866, + 0.004427300300449133, + 0.00444827601313591, + 0.004525273572653532, + 0.005442049354314804, + 0.006777444388717413, + 0.0021159490570425987, + 0.001185635570436716, + 0.0027976033743470907, + 0.003592920722439885, + 0.002256891457363963, + 0.002238362329080701, + 0.004155686590820551, + 0.004200059454888105, + 0.003061166498810053, + 0.001985460054129362, + 0.006675136741250753, + 0.009336058981716633, + 0.0029010886792093515, + 0.0018195494776591659, + 0.005321556702256203, + 0.004669870715588331, + 0.005494709126651287, + 0.007440473418682814, + 0.00753368902951479, + 0.00576488533988595, + 0.01866581104695797, + 0.007224113680422306, + 0.008987625129520893, + 0.010897818952798843, + 0.005572603549808264, + 0.008012106642127037, + 0.005265393294394016, + 0.005357825662940741, + 0.009902529418468475, + 0.022327279672026634, + 0.008456263691186905, + 0.008668312802910805, + 0.005181277170777321, + 0.006483347620815039, + 0.0029227719642221928, + 0.00506600271910429, + 0.0049302284605801105, + 0.0067672994919121265, + 0.005837974604219198, + 0.006621099077165127, + 0.006902576424181461, + 0.006626080255955458, + 0.00630028173327446, + 0.004156959243118763, + 0.005589309148490429, + 0.008604797534644604, + 0.014733457937836647, + 0.0077001056633889675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.10030201822519302 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02632831409573555 + ], + "opposite_logprob_tokens": [ + -3.6513283252716064 + ], + "logprob_reward": -0.02632831409573555, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02632831409573555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17398691177368164, + 0.08042827993631363, + 0.04103908687829971, + 0.07381705194711685, + 0.055780839174985886, + 0.06343479454517365, + 0.13686779141426086, + 0.158029705286026, + 0.044109996408224106, + 0.17250551283359528 + ], + "credit_map": [ + 0.019181353971362114, + 0.014744400978088379, + 0.03653029724955559, + 0.02818116545677185, + 0.025848979130387306, + 0.014469973742961884, + 0.021379444748163223, + 0.013651282526552677, + 0.007419339381158352, + 0.008325416594743729, + 0.020656688138842583, + 0.016820799559354782, + 0.006411897949874401, + 0.0045358664356172085, + 0.009959298186004162, + 0.006298976484686136, + 0.0039498754777014256, + 0.004167310893535614, + 0.004695242270827293, + 0.005919994320720434, + 0.004656508564949036, + 0.004999910015612841, + 0.006190262734889984, + 0.006459981203079224, + 0.004956671502441168, + 0.007311748340725899, + 0.01849033124744892, + 0.0098074059933424, + 0.0069260806776583195, + 0.006926032714545727, + 0.01088783796876669, + 0.008510936982929707, + 0.004580023232847452, + 0.004309903830289841, + 0.009846724569797516, + 0.01112106628715992, + 0.0055110864341259, + 0.005143825896084309, + 0.008275249041616917, + 0.006992956157773733, + 0.005056031513959169, + 0.0072160824202001095, + 0.007235644850879908, + 0.007527953013777733, + 0.016059357672929764, + 0.007309658452868462, + 0.005136646796017885, + 0.007893412373960018, + 0.011385513469576836, + 0.019354507327079773, + 0.004906493704766035, + 0.0061189765110611916, + 0.04943216219544411, + 0.015405401587486267, + 0.009928476065397263, + 0.020336264744400978, + 0.00785831082612276, + 0.016351809725165367, + 0.009403754025697708, + 0.012776456773281097, + 0.02296537160873413, + 0.054043449461460114, + 0.014521988108754158, + 0.020108552649617195, + 0.003191995667293668, + 0.0037454182747751474, + 0.00411471351981163, + 0.009868004359304905, + 0.0035219937562942505, + 0.0043638055212795734, + 0.008203340694308281, + 0.007100721355527639, + 0.01284094899892807, + 0.01042872853577137, + 0.017824120819568634, + 0.012817902490496635, + 0.013938994146883488, + 0.0401463657617569, + 0.039728134870529175, + 0.024780312553048134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02632831409573555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a heavy nighttime snowfall, a high-angle, wide-angle shot slowly sweeps left across a vast, floodlit football stadium. The camera captures the entire spectacle, showing a field almost entirely covered in pristine white snow, with barely a hint of green turf beneath. Thick, swirling snowflakes drift relentlessly down, caught by the powerful, cold white light from the overhead stadium lights, casting long, dramatic shadows across the field. The tribune is packed with a dense, roaring crowd, crammed with spectators clad in heavy winter coats, colorful woolen hats, and team scarves. Their collective breath hangs like a visible mist in the chilly air. On the field below, players, clad in vibrant, solid jerseys—one team all-red, the other all-blue—move with energy and determination, vying for a bright orange ball. They sprint, slide, and tackle, kicking up explosive plumes of snow with every powerful stride and challenge, creating a scene of intense, raw athletic passion that contrasts with the harsh winter weather and creates a thrilling atmosphere.", + "seed": 6272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6272_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": -0.06252744346857071, + "vlm_logprob": -0.06252744346857071, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06252744346857071, + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0389007069170475 + ], + "opposite_logprob_tokens": [ + -3.41390061378479 + ], + "logprob_reward": -0.0389007069170475, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0389007069170475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20040927827358246, + 0.18228377401828766, + 0.01809200458228588, + 0.05701480433344841, + 0.05518915504217148, + 0.30595916509628296, + 0.018016712740063667, + 0.010289880447089672, + 0.14271856844425201, + 0.010026669129729271 + ], + "credit_map": [ + 0.012435771524906158, + 0.061786193400621414, + 0.009883354417979717, + 0.010525383986532688, + 0.07134047895669937, + 0.016521066427230835, + 0.007752612233161926, + 0.010164408944547176, + 0.009545943699777126, + 0.05539895221590996, + 0.010897986590862274, + 0.00674727326259017, + 0.06807040423154831, + 0.011949599720537663, + 0.011842865496873856, + 0.007830738089978695, + 0.0012065587798133492, + 0.002106734551489353, + 0.00180023442953825, + 0.0026911075692623854, + 0.0033553102985024452, + 0.001297198235988617, + 0.003501880681142211, + 0.0021329785231500864, + 0.0031558824703097343, + 0.020964832976460457, + 0.00207921932451427, + 0.002094787545502186, + 0.02009359933435917, + 0.002765991957858205, + 0.002436918206512928, + 0.0034235725179314613, + 0.0046612415462732315, + 0.013872670009732246, + 0.005902130156755447, + 0.004683211911469698, + 0.00854596495628357, + 0.0051832739263772964, + 0.006428261287510395, + 0.005912397988140583, + 0.02043505571782589, + 0.12837567925453186, + 0.004839728120714426, + 0.005569937638938427, + 0.11805914342403412, + 0.015328029170632362, + 0.00599819328635931, + 0.007353365886956453, + 0.0010567426215857267, + 0.004473804961889982, + 0.0026489021256566048, + 0.0012263529933989048, + 0.004500711336731911, + 0.0013823335757479072, + 0.0012906081974506378, + 0.0014372550649568439, + 0.0007656087982468307, + 0.001161151914857328, + 0.0013861621264368296, + 0.0014435204211622477, + 0.0013120454968884587, + 0.0016800236189737916, + 0.0012702109524980187, + 0.0012711567105725408, + 0.008761337026953697, + 0.04041222110390663, + 0.006187158636748791, + 0.012397256679832935, + 0.036631107330322266, + 0.010976236313581467, + 0.015822695568203926, + 0.011530548334121704, + 0.0008872740436345339, + 0.0026854639872908592, + 0.0010223651770502329, + 0.000787908211350441, + 0.0015371614135801792, + 0.0013215737417340279, + 0.0008629736257717013, + 0.0009219490457326174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0389007069170475 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293032020330429 + ], + "opposite_logprob_tokens": [ + -2.2379302978515625 + ], + "logprob_reward": -0.11293032020330429, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293032020330429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07483214139938354, + 0.03252212330698967, + 0.04869932681322098, + 0.025615647435188293, + 0.027696965262293816, + 0.03975439816713333, + 0.06339526176452637, + 0.048746801912784576, + 0.6293248534202576, + 0.009412596933543682 + ], + "credit_map": [ + 0.008747194893658161, + 0.010658579878509045, + 0.011681499890983105, + 0.01042737066745758, + 0.008114667609333992, + 0.008575876243412495, + 0.008965290151536465, + 0.007661645766347647, + 0.00342188635841012, + 0.0074853440746665, + 0.0030602740589529276, + 0.0022835053969174623, + 0.006189382169395685, + 0.005017559044063091, + 0.0025593906175345182, + 0.0025047780945897102, + 0.004680002573877573, + 0.011884779669344425, + 0.005144993308931589, + 0.005941907875239849, + 0.006600663531571627, + 0.003046230645850301, + 0.005179962143301964, + 0.0062207807786762714, + 0.0027005921583622694, + 0.0029505249112844467, + 0.0029271889943629503, + 0.0026501582469791174, + 0.0030131672974675894, + 0.004351482260972261, + 0.0033288891427218914, + 0.0036936413962394, + 0.0022208737209439278, + 0.0060432711616158485, + 0.002432974288240075, + 0.003761415369808674, + 0.004429790657013655, + 0.002484285971149802, + 0.003094105515629053, + 0.003230244852602482, + 0.0030338773503899574, + 0.012860588729381561, + 0.002460250398144126, + 0.0023733898997306824, + 0.010532031767070293, + 0.0032057571224868298, + 0.0029803935904055834, + 0.0023081046529114246, + 0.0039231786504387856, + 0.02076491340994835, + 0.0037292749620974064, + 0.0032883828971534967, + 0.021324653178453445, + 0.003824755782261491, + 0.0036094749812036753, + 0.0029306216165423393, + 0.003896943759173155, + 0.008136219345033169, + 0.005982831586152315, + 0.007414765190333128, + 0.0069951461628079414, + 0.004511888604611158, + 0.005523245315998793, + 0.006285754032433033, + 0.034192826598882675, + 0.20067231357097626, + 0.024969659745693207, + 0.04250948131084442, + 0.1937144696712494, + 0.03936946392059326, + 0.05736832693219185, + 0.03652820736169815, + 0.0008430295274592936, + 0.001181884203106165, + 0.0011942529818043113, + 0.0011544176377356052, + 0.0009864239254966378, + 0.0017749735852703452, + 0.0012087106006219983, + 0.0010689024347811937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.11293032020330429 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784713387489319 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784713387489319, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784713387489319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10133623331785202, + 0.06978911906480789, + 0.044946733862161636, + 0.07097652554512024, + 0.024137012660503387, + 0.334175705909729, + 0.12667876482009888, + 0.017082413658499718, + 0.19713258743286133, + 0.01374484971165657 + ], + "credit_map": [ + 0.011648566462099552, + 0.01935875043272972, + 0.011149267666041851, + 0.01036251150071621, + 0.018177444115281105, + 0.013041791506111622, + 0.008155813440680504, + 0.009442097507417202, + 0.004649633541703224, + 0.017673619091510773, + 0.0060029104351997375, + 0.003907922655344009, + 0.0234824251383543, + 0.004998491611331701, + 0.005143571179360151, + 0.003930552862584591, + 0.0044789994135499, + 0.006290454417467117, + 0.004321104846894741, + 0.005563134327530861, + 0.005428864620625973, + 0.004278075881302357, + 0.008653712458908558, + 0.00593239301815629, + 0.004192269407212734, + 0.02782604843378067, + 0.002370412228628993, + 0.002357339020818472, + 0.026113996282219887, + 0.0030882572755217552, + 0.002678924473002553, + 0.0023492840118706226, + 0.001847655512392521, + 0.004885753616690636, + 0.002518032444640994, + 0.003353187581524253, + 0.0033004169818013906, + 0.002268324140459299, + 0.003119255183264613, + 0.002844390692189336, + 0.02236368879675865, + 0.14118395745754242, + 0.004353903699666262, + 0.0060885995626449585, + 0.12952065467834473, + 0.016803791746497154, + 0.006225159391760826, + 0.0076359775848686695, + 0.008080830797553062, + 0.05046207457780838, + 0.005194989964365959, + 0.0027539979200810194, + 0.04638952016830444, + 0.006812255829572678, + 0.0035415205638855696, + 0.0034435829147696495, + 0.0013665402075275779, + 0.0029904968105256557, + 0.0019806812051683664, + 0.002789711579680443, + 0.002134511945769191, + 0.002937713172286749, + 0.00138479087036103, + 0.0014979694969952106, + 0.010235127061605453, + 0.05875621736049652, + 0.010233148001134396, + 0.01530752144753933, + 0.057620927691459656, + 0.012095034122467041, + 0.01890992559492588, + 0.013974694535136223, + 0.0013535416219383478, + 0.0026936575304716825, + 0.0012027673656120896, + 0.001172723714262247, + 0.0017291855765506625, + 0.0018667153781279922, + 0.0016699183033779263, + 0.0020563420839607716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.009784713387489319 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294001340866089 + ], + "opposite_logprob_tokens": [ + -2.2379400730133057 + ], + "logprob_reward": -0.11294001340866089, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294001340866089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2121783196926117, + 0.0872606709599495, + 0.041786327958106995, + 0.10175507515668869, + 0.016417264938354492, + 0.030347831547260284, + 0.12149020284414291, + 0.020638318732380867, + 0.34921666979789734, + 0.018909290432929993 + ], + "credit_map": [ + 0.011705231852829456, + 0.051541805267333984, + 0.026455186307430267, + 0.015659131109714508, + 0.06840300559997559, + 0.014140413142740726, + 0.01216831710189581, + 0.01210523396730423, + 0.006088377442210913, + 0.02188521809875965, + 0.006920628249645233, + 0.006338362582027912, + 0.02911234274506569, + 0.007289614994078875, + 0.005105795804411173, + 0.004520328249782324, + 0.0031250889878720045, + 0.010941440239548683, + 0.0028524508234113455, + 0.003642773488536477, + 0.009977404959499836, + 0.0022739595733582973, + 0.005958962254226208, + 0.0030142467003315687, + 0.005747129675000906, + 0.03043873980641365, + 0.0057612513191998005, + 0.006085994653403759, + 0.031872499734163284, + 0.00872532743960619, + 0.005825331434607506, + 0.0072987996973097324, + 0.0012842559954151511, + 0.003452501492574811, + 0.0016780153382569551, + 0.0022925303783267736, + 0.002524162409827113, + 0.0009655113099142909, + 0.0018607265083119273, + 0.002359562087804079, + 0.0017488034209236503, + 0.00909545086324215, + 0.0022659539245069027, + 0.0027678930200636387, + 0.007922385819256306, + 0.0024907065089792013, + 0.0020106120500713587, + 0.002046026522293687, + 0.007484799716621637, + 0.045202307403087616, + 0.006570505443960428, + 0.004138879012316465, + 0.04197866842150688, + 0.006807469297200441, + 0.005052302964031696, + 0.004255266394466162, + 0.002078474499285221, + 0.0026398771442472935, + 0.002546158619225025, + 0.0038800823967903852, + 0.0023346785455942154, + 0.0022604744881391525, + 0.0024374593049287796, + 0.002461113501340151, + 0.019101949408650398, + 0.11377651989459991, + 0.01391823124140501, + 0.021909238770604134, + 0.10951117426156998, + 0.021760983392596245, + 0.029767870903015137, + 0.019470710307359695, + 0.0023672045208513737, + 0.0018941264133900404, + 0.0025626940187066793, + 0.0018309734296053648, + 0.0021922693122178316, + 0.0040811519138514996, + 0.0017450375016778708, + 0.0022358326241374016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.11294001340866089 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808146342635155 + ], + "opposite_logprob_tokens": [ + -3.288081407546997 + ], + "logprob_reward": -0.03808146342635155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808146342635155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20406685769557953, + 0.049001626670360565, + 0.0329543836414814, + 0.0316673181951046, + 0.04661984741687775, + 0.41895610094070435, + 0.06694551557302475, + 0.05829795449972153, + 0.07449744641780853, + 0.016992952674627304 + ], + "credit_map": [ + 0.016018835827708244, + 0.03240688517689705, + 0.02987338975071907, + 0.022352658212184906, + 0.04146544262766838, + 0.025472937151789665, + 0.01741148717701435, + 0.01906522549688816, + 0.003784644417464733, + 0.011110648512840271, + 0.008002149872481823, + 0.004106481093913317, + 0.010189104825258255, + 0.0031213085167109966, + 0.00505185266956687, + 0.003635440254583955, + 0.002738027134910226, + 0.0024540408048778772, + 0.004181175492703915, + 0.006882976274937391, + 0.0036863358691334724, + 0.0021994563285261393, + 0.006558123044669628, + 0.004254248924553394, + 0.0018337296787649393, + 0.00640068668872118, + 0.0029473896138370037, + 0.003243587911128998, + 0.006570545956492424, + 0.003493616357445717, + 0.003463657107204199, + 0.003714107209816575, + 0.003912337590008974, + 0.00961116049438715, + 0.004621350672096014, + 0.005147236865013838, + 0.005065117496997118, + 0.005856047384440899, + 0.0054899416863918304, + 0.0069166552275419235, + 0.028193462640047073, + 0.17686063051223755, + 0.005619864445179701, + 0.00760611891746521, + 0.1621958464384079, + 0.021003294736146927, + 0.00791312288492918, + 0.00956378411501646, + 0.00379022303968668, + 0.022700147703289986, + 0.004001901485025883, + 0.0025964488741010427, + 0.021871687844395638, + 0.004090654663741589, + 0.00450432812795043, + 0.003390128258615732, + 0.003811192000284791, + 0.005203184671700001, + 0.009849912486970425, + 0.011953650042414665, + 0.007502870634198189, + 0.004544415976852179, + 0.0070536695420742035, + 0.008379059843719006, + 0.004303697030991316, + 0.02127036824822426, + 0.0037482476327568293, + 0.005944308824837208, + 0.020006082952022552, + 0.004744159989058971, + 0.008611421100795269, + 0.005869156681001186, + 0.0018423624569550157, + 0.001934248604811728, + 0.0026085982099175453, + 0.0015278676291927695, + 0.001758367521688342, + 0.0033595238346606493, + 0.0021349529270082712, + 0.001827032072469592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.03808146342635155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Amidst a heavy nighttime snowfall, a high-angle, wide-angle shot slowly sweeps left across a vast, floodlit football stadium. The camera captures the entire spectacle, showing a field almost entirely covered in pristine white snow, with barely a hint of green turf beneath. Thick, swirling snowflakes drift relentlessly down, caught by the powerful, cold white light from the overhead stadium lights, casting long, dramatic shadows across the field. The tribune is packed with a dense, roaring crowd, crammed with spectators clad in heavy winter coats, colorful woolen hats, and team scarves. Their collective breath hangs like a visible mist in the chilly air. On the field below, players, clad in vibrant, solid jerseys—one team all-red, the other all-blue—move with energy and determination, vying for a bright orange ball. They sprint, slide, and tackle, kicking up explosive plumes of snow with every powerful stride and challenge, creating a scene of intense, raw athletic passion that contrasts with the harsh winter weather and creates a thrilling atmosphere.", + "seed": 6273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_1/index_9275_seed_6273_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": -0.0614502239972353, + "vlm_logprob": -0.0614502239972353, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.0614502239972353, + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0935041755437851 + ], + "opposite_logprob_tokens": [ + -2.4685041904449463 + ], + "logprob_reward": -0.0935041755437851, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0935041755437851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11468210816383362, + 0.1464337259531021, + 0.14645104110240936, + 0.08219349384307861, + 0.15278615057468414, + 0.24058741331100464, + 0.024143293499946594, + 0.016132375225424767, + 0.024619832634925842, + 0.051970597356557846 + ], + "credit_map": [ + 0.014117207378149033, + 0.0278056338429451, + 0.008269487880170345, + 0.008166059851646423, + 0.013221080414950848, + 0.015803290531039238, + 0.009765398688614368, + 0.01753394491970539, + 0.011651637963950634, + 0.023357117548584938, + 0.018579505383968353, + 0.01664791814982891, + 0.016493091359734535, + 0.021943097934126854, + 0.017081674188375473, + 0.020679671317338943, + 0.014498895034193993, + 0.025179114192724228, + 0.011093382723629475, + 0.013264661654829979, + 0.028385240584611893, + 0.017471525818109512, + 0.01597934029996395, + 0.02057887241244316, + 0.009682374075055122, + 0.012779800221323967, + 0.007487100549042225, + 0.009232793934643269, + 0.014093113131821156, + 0.010722251608967781, + 0.007248061243444681, + 0.010947993956506252, + 0.016718605533242226, + 0.023556362837553024, + 0.009405731223523617, + 0.01653221994638443, + 0.031780097633600235, + 0.026235103607177734, + 0.011444355361163616, + 0.017113657668232918, + 0.032174449414014816, + 0.043611831963062286, + 0.011893648654222488, + 0.039004847407341, + 0.048703789710998535, + 0.02665531076490879, + 0.02377004735171795, + 0.014773465692996979, + 0.004927138797938824, + 0.002626139670610428, + 0.002087715547531843, + 0.002409275621175766, + 0.003921409137547016, + 0.003338242182508111, + 0.002257655607536435, + 0.002575714373961091, + 0.003393104998394847, + 0.001585576101206243, + 0.002326969290152192, + 0.0018538163276389241, + 0.0021123457700014114, + 0.0020386071410030127, + 0.0015903400490060449, + 0.0012316156644374132, + 0.004859980195760727, + 0.0016472104471176863, + 0.003317761467769742, + 0.003469117684289813, + 0.0015313798794522882, + 0.002866351278498769, + 0.0035040564835071564, + 0.003423972986638546, + 0.009004451334476471, + 0.0072119832038879395, + 0.006047704257071018, + 0.005107032600790262, + 0.009301911108195782, + 0.006264733150601387, + 0.0040781619027256966, + 0.004954618401825428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.0935041755437851 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495745688676834 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495745688676834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495745688676834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10996246337890625, + 0.04319640249013901, + 0.1281593143939972, + 0.04642583802342415, + 0.07736486941576004, + 0.25454097986221313, + 0.02564089745283127, + 0.02366957999765873, + 0.23085951805114746, + 0.06018010899424553 + ], + "credit_map": [ + 0.0074305227026343346, + 0.009830201044678688, + 0.016804412007331848, + 0.015308876521885395, + 0.009412608109414577, + 0.013950489461421967, + 0.010392612777650356, + 0.02683275006711483, + 0.004383707884699106, + 0.007242987863719463, + 0.005883374717086554, + 0.006151655688881874, + 0.004089648369699717, + 0.004806086886674166, + 0.005428879987448454, + 0.0052100662142038345, + 0.011298749595880508, + 0.01860462687909603, + 0.013356250710785389, + 0.014991797506809235, + 0.020606456324458122, + 0.014305499382317066, + 0.01710602641105652, + 0.017889922484755516, + 0.004898222628980875, + 0.00770824309438467, + 0.00373562122695148, + 0.004398007411509752, + 0.007382197305560112, + 0.005208518356084824, + 0.006493574008345604, + 0.006601456552743912, + 0.009143490344285965, + 0.011687548831105232, + 0.005554624367505312, + 0.009644796140491962, + 0.015115327201783657, + 0.011209913529455662, + 0.006363296881318092, + 0.008645872585475445, + 0.032638028264045715, + 0.0431920550763607, + 0.013679456897079945, + 0.04162704944610596, + 0.049259334802627563, + 0.0271513219922781, + 0.027489744126796722, + 0.019503995776176453, + 0.0045852260664105415, + 0.0025843956973403692, + 0.002982331672683358, + 0.0032461024820804596, + 0.002764900214970112, + 0.0038300552405416965, + 0.0032693224493414164, + 0.0023785652592778206, + 0.0038051940500736237, + 0.002366482513025403, + 0.004055347293615341, + 0.0028134111780673265, + 0.002659338526427746, + 0.002864251844584942, + 0.0028735012747347355, + 0.002232055179774761, + 0.051509782671928406, + 0.015779811888933182, + 0.023175297304987907, + 0.03535958379507065, + 0.01495073176920414, + 0.029495546594262123, + 0.021737784147262573, + 0.03885100409388542, + 0.010798855684697628, + 0.00793683622032404, + 0.007586668245494366, + 0.007418734021484852, + 0.008359771221876144, + 0.0073653096333146095, + 0.004601770546287298, + 0.006112166680395603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.05495745688676834 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14043238759040833, + 0.2429325133562088, + 0.0897352322936058, + 0.053420212119817734, + 0.08846072107553482, + 0.053980130702257156, + 0.048832133412361145, + 0.034995097666978836, + 0.07026698440313339, + 0.17694459855556488 + ], + "credit_map": [ + 0.012566978111863136, + 0.015661124140024185, + 0.018571004271507263, + 0.01685141772031784, + 0.02136293798685074, + 0.02106153592467308, + 0.015613429248332977, + 0.01874396577477455, + 0.027345621958374977, + 0.04220392927527428, + 0.036972399801015854, + 0.030987324193120003, + 0.026993505656719208, + 0.027549920603632927, + 0.028554461896419525, + 0.02232535555958748, + 0.008595693856477737, + 0.016850853338837624, + 0.006026472430676222, + 0.009960410185158253, + 0.014417914673686028, + 0.00985258910804987, + 0.011142546311020851, + 0.012888758443295956, + 0.005678094457834959, + 0.007292945869266987, + 0.004773058462888002, + 0.006724510341882706, + 0.006580780725926161, + 0.005743826273828745, + 0.008558091707527637, + 0.008068910799920559, + 0.008114300668239594, + 0.011210109107196331, + 0.010589762590825558, + 0.012171397916972637, + 0.01687232032418251, + 0.014207294210791588, + 0.007425898220390081, + 0.007869642227888107, + 0.009191569872200489, + 0.0063940794207155704, + 0.006983507890254259, + 0.006139415316283703, + 0.009327854961156845, + 0.006850111298263073, + 0.004109819885343313, + 0.004983780439943075, + 0.008869834244251251, + 0.005253099836409092, + 0.0048246621154248714, + 0.005871253088116646, + 0.006946714129298925, + 0.0064389524050056934, + 0.005973479710519314, + 0.004654145799577236, + 0.008348100818693638, + 0.0021545295603573322, + 0.00511575024574995, + 0.004257270134985447, + 0.002326823305338621, + 0.004775891080498695, + 0.003580095712095499, + 0.004436640068888664, + 0.01099494006484747, + 0.006569806486368179, + 0.00909656286239624, + 0.009982744231820107, + 0.0068876855075359344, + 0.007253700867295265, + 0.008759452030062675, + 0.010722094215452671, + 0.02823706530034542, + 0.021403351798653603, + 0.021565603092312813, + 0.01618625782430172, + 0.02924274653196335, + 0.02055627480149269, + 0.01565588265657425, + 0.024097450077533722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.016070127487182617 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293735355138779 + ], + "opposite_logprob_tokens": [ + -2.2379374504089355 + ], + "logprob_reward": -0.11293735355138779, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293735355138779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1598205268383026, + 0.3244296908378601, + 0.04671710729598999, + 0.03731418773531914, + 0.04240599647164345, + 0.16001290082931519, + 0.048108622431755066, + 0.030258072540163994, + 0.06668780744075775, + 0.08424511551856995 + ], + "credit_map": [ + 0.012978804297745228, + 0.020871438086032867, + 0.014813294634222984, + 0.012798111885786057, + 0.015614699572324753, + 0.01929437555372715, + 0.01879754289984703, + 0.04465225711464882, + 0.03174484521150589, + 0.053549621254205704, + 0.05276346951723099, + 0.04832432046532631, + 0.03289651870727539, + 0.03548445552587509, + 0.041631150990724564, + 0.028035297989845276, + 0.004461830947548151, + 0.007446424104273319, + 0.0036592530086636543, + 0.0038315230049192905, + 0.008869856595993042, + 0.005734019447118044, + 0.005597716197371483, + 0.007116486318409443, + 0.0039551821537315845, + 0.00407921290025115, + 0.004209305625408888, + 0.005744110327214003, + 0.004809495527297258, + 0.004742125049233437, + 0.004810705780982971, + 0.004964049905538559, + 0.004301594104617834, + 0.0033499845303595066, + 0.004959973506629467, + 0.009242168627679348, + 0.005115622654557228, + 0.004819174762815237, + 0.0048352633602917194, + 0.005782212596386671, + 0.019125722348690033, + 0.02763272263109684, + 0.01029924489557743, + 0.027413586154580116, + 0.029948970302939415, + 0.016759086400270462, + 0.01691613346338272, + 0.011917428113520145, + 0.010431402362883091, + 0.0036608967930078506, + 0.005553656257688999, + 0.006110569462180138, + 0.004899780731648207, + 0.006377412471920252, + 0.00680227717384696, + 0.004272628575563431, + 0.007764099165797234, + 0.0021787898149341345, + 0.003201138461008668, + 0.0034311639610677958, + 0.003015858121216297, + 0.004414469003677368, + 0.0031509771943092346, + 0.0031015779823064804, + 0.009997587651014328, + 0.004729902371764183, + 0.0082220658659935, + 0.011697104200720787, + 0.004877794533967972, + 0.006369137205183506, + 0.007514408323913813, + 0.013279808685183525, + 0.017045965418219566, + 0.008679457008838654, + 0.010965538211166859, + 0.012933130376040936, + 0.008207231760025024, + 0.009189504198729992, + 0.007776742801070213, + 0.00944754108786583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.11293735355138779 + }, + { + "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 football stadium during heavy nighttime snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot high-angle, wide-angle, and sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the field covered in white snow with barely visible green turf, illuminated by powerful overhead stadium lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are players visible wearing solid all-red and all-blue jerseys, vying for a bright orange ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tribune packed with a dense crowd wearing heavy winter coats, colorful woolen hats, and team scarves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978200651705265 + ], + "opposite_logprob_tokens": [ + -3.5297820568084717 + ], + "logprob_reward": -0.02978200651705265, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978200651705265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394912153482437, + 0.4587239921092987, + 0.026550965383648872, + 0.02705121599137783, + 0.033098235726356506, + 0.1804303675889969, + 0.014588595367968082, + 0.010798181407153606, + 0.08374536782503128, + 0.025521913543343544 + ], + "credit_map": [ + 0.011181852780282497, + 0.01319396123290062, + 0.028240976855158806, + 0.01677376963198185, + 0.01208529993891716, + 0.014105139300227165, + 0.016727346926927567, + 0.027182862162590027, + 0.04472624510526657, + 0.07340136170387268, + 0.06621304899454117, + 0.062216617166996, + 0.04947929456830025, + 0.06093516945838928, + 0.05323674529790878, + 0.04851548746228218, + 0.0019800658337771893, + 0.0033861116971820593, + 0.002654984127730131, + 0.0034219161607325077, + 0.003944604191929102, + 0.0032213761005550623, + 0.0036203109193593264, + 0.004321596585214138, + 0.0031185296829789877, + 0.0030013923533260822, + 0.0032001500949263573, + 0.00354556436650455, + 0.0034917842131108046, + 0.0036235693842172623, + 0.003187586087733507, + 0.0038826395757496357, + 0.0033985411282628775, + 0.004157481715083122, + 0.0037877399008721113, + 0.004768554121255875, + 0.005467990878969431, + 0.0038313039112836123, + 0.0037041448522359133, + 0.003982477355748415, + 0.02201060578227043, + 0.032440681010484695, + 0.009439374320209026, + 0.0300525464117527, + 0.035905465483665466, + 0.018986931070685387, + 0.01928228698670864, + 0.012312485836446285, + 0.001509913825429976, + 0.0009867026237770915, + 0.0029499514494091272, + 0.002214150270447135, + 0.0012504232581704855, + 0.0016078276094049215, + 0.0027294144965708256, + 0.0013402111362665892, + 0.0017658044816926122, + 0.0008734787115827203, + 0.0022534355521202087, + 0.0012250661384314299, + 0.0009760208777152002, + 0.001478905905969441, + 0.0013708064798265696, + 0.0008546633762307465, + 0.0131353959441185, + 0.0047728270292282104, + 0.011929349042475224, + 0.014215642586350441, + 0.004874385427683592, + 0.009338553994894028, + 0.010632556863129139, + 0.014846651814877987, + 0.003584527876228094, + 0.0026021709199994802, + 0.0032793013378977776, + 0.003998949658125639, + 0.0024048888590186834, + 0.0024399799294769764, + 0.002809327794238925, + 0.004402767866849899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.02978200651705265 + } + ], + "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/0000060/group_10/index_tensor([126, 126])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_tensor([126, 126])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b6567829c545c5b037b560541529ecfb7f7019b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_tensor([126, 126])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle shot tilts steadily upward, its movement synchronized with the powerful rise of a single, massive blade from a wind turbine. Made of sleek gray-white fiberglass, the blade fills the frame, its smooth, aerodynamic surface capturing the first golden rays of the rising sun, forming a brilliant leading edge against the deep blue of the clear morning sky. As it reaches its peak, the camera continues to follow its relentless, hypnotic rotation, cutting cleanly through the air with powerful grace, its tip slightly blurred by sheer speed. Below, rolling green hills, vibrant and lush, stretch to the horizon, while a gentle morning mist gathers softly in the valley. The scene is bathed in the cool, clear light of dawn, creating long, soft shadows and evoking an atmosphere of majestic power and tranquil, clean energy.", + "seed": 20834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20834_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": -0.15683176461607218, + "vlm_logprob": -0.15683176461607218, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.15683176461607218, + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.58104008436203 + ], + "opposite_logprob_tokens": [ + -0.83104008436203 + ], + "logprob_reward": -0.58104008436203, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.58104008436203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11336012929677963, + 0.030071722343564034, + 0.1995389461517334, + 0.111473947763443, + 0.05078079178929329, + 0.039576124399900436, + 0.13755251467227936, + 0.05482315272092819, + 0.08154480159282684, + 0.18127785623073578 + ], + "credit_map": [ + 0.00743876164779067, + 0.012155849486589432, + 0.0185175109654665, + 0.030764535069465637, + 0.013812508434057236, + 0.010375783778727055, + 0.010692917741835117, + 0.009602255187928677, + 0.005508792586624622, + 0.0036535789258778095, + 0.0038212635554373264, + 0.002287389012053609, + 0.004510985687375069, + 0.0027322128880769014, + 0.0032278525177389383, + 0.004329645074903965, + 0.031241899356245995, + 0.03210299834609032, + 0.022951224818825722, + 0.018434401601552963, + 0.0298449844121933, + 0.023625483736395836, + 0.015990247949957848, + 0.025347691029310226, + 0.01073923334479332, + 0.013772628270089626, + 0.005413773003965616, + 0.005511647555977106, + 0.0380694642663002, + 0.014491402544081211, + 0.01422464195638895, + 0.009251158684492111, + 0.00541273970156908, + 0.007736963219940662, + 0.004381151869893074, + 0.004288214258849621, + 0.012111099436879158, + 0.006582984700798988, + 0.005071687512099743, + 0.005195955745875835, + 0.0034522833302617073, + 0.0095670809969306, + 0.0026647222694009542, + 0.002780241658911109, + 0.005769398063421249, + 0.009205952286720276, + 0.0030586463399231434, + 0.003077798057347536, + 0.013186885975301266, + 0.05050373449921608, + 0.00578880961984396, + 0.0062614865601062775, + 0.010506317019462585, + 0.040469761937856674, + 0.003672598395496607, + 0.007162929978221655, + 0.005955993197858334, + 0.01258790958672762, + 0.0035735478159040213, + 0.0038992997724562883, + 0.010257652960717678, + 0.00854365061968565, + 0.0059450538828969, + 0.004060045816004276, + 0.009499836713075638, + 0.020585406571626663, + 0.007090600673109293, + 0.0057290093973279, + 0.013993619941174984, + 0.010457449592649937, + 0.009680761955678463, + 0.004508113022893667, + 0.011999793350696564, + 0.05126594752073288, + 0.01185593567788601, + 0.03141104057431221, + 0.013713928870856762, + 0.04464181512594223, + 0.00971178337931633, + 0.006677613593637943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.58104008436203 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004282623529434 + ], + "opposite_logprob_tokens": [ + -2.695042848587036 + ], + "logprob_reward": -0.07004282623529434, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004282623529434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08790453523397446, + 0.02773180417716503, + 0.10724768787622452, + 0.06336595863103867, + 0.07745511084794998, + 0.04801429435610771, + 0.38398605585098267, + 0.021552007645368576, + 0.06491243839263916, + 0.11783016473054886 + ], + "credit_map": [ + 0.008052568882703781, + 0.011948106810450554, + 0.010435669682919979, + 0.01985788531601429, + 0.012283760122954845, + 0.011003216728568077, + 0.006507223006337881, + 0.007816108874976635, + 0.004520747810602188, + 0.0033070009667426348, + 0.0036771332379430532, + 0.004017566330730915, + 0.002486886689439416, + 0.002332596341148019, + 0.002424051985144615, + 0.004965822212398052, + 0.01798076368868351, + 0.020706959068775177, + 0.01161160133779049, + 0.008897995576262474, + 0.012677323073148727, + 0.014942201785743237, + 0.007543311454355717, + 0.012887522578239441, + 0.007509516552090645, + 0.007142559625208378, + 0.006022321060299873, + 0.005509242881089449, + 0.018002651631832123, + 0.006534383166581392, + 0.006111628841608763, + 0.006533648353070021, + 0.008927443064749241, + 0.011867230758070946, + 0.008928737603127956, + 0.007687801960855722, + 0.013886884786188602, + 0.010359230451285839, + 0.0073501490987837315, + 0.008447629399597645, + 0.005192473065108061, + 0.005539314821362495, + 0.004963702522218227, + 0.005003779195249081, + 0.011407754383981228, + 0.004773896653205156, + 0.005910371895879507, + 0.005222998093813658, + 0.028410455211997032, + 0.15249402821063995, + 0.01645672507584095, + 0.01574314385652542, + 0.020006394013762474, + 0.1354132890701294, + 0.005162831395864487, + 0.010299189016222954, + 0.002259120112285018, + 0.0037988668773323298, + 0.0022338510025292635, + 0.0016916025197133422, + 0.0042664106003940105, + 0.0025113453157246113, + 0.002387132728472352, + 0.0024036788381636143, + 0.004966646432876587, + 0.008915236219763756, + 0.007094907574355602, + 0.006190832704305649, + 0.00919413659721613, + 0.010421300306916237, + 0.009708224795758724, + 0.008421159349381924, + 0.010155429132282734, + 0.04310326650738716, + 0.006445649079978466, + 0.011374225839972496, + 0.005638505797833204, + 0.03396580368280411, + 0.003249955363571644, + 0.0038973409682512283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.07004282623529434 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16425149142742157, + 0.05011064186692238, + 0.207172691822052, + 0.047668639570474625, + 0.021080177277326584, + 0.038161102682352066, + 0.2776506245136261, + 0.03131270781159401, + 0.0828380212187767, + 0.0797538235783577 + ], + "credit_map": [ + 0.013956278562545776, + 0.01517409272491932, + 0.0375920832157135, + 0.04505237936973572, + 0.01603897474706173, + 0.009440816938877106, + 0.01163867861032486, + 0.015358193777501583, + 0.006512906402349472, + 0.007224618922919035, + 0.007580724079161882, + 0.006154391448944807, + 0.006001112516969442, + 0.004089802969247103, + 0.004164753947407007, + 0.008382333442568779, + 0.033835962414741516, + 0.03759331256151199, + 0.023097075521945953, + 0.016690298914909363, + 0.025271698832511902, + 0.029979020357131958, + 0.015256972052156925, + 0.025448372587561607, + 0.004767946433275938, + 0.005598447285592556, + 0.006289483513683081, + 0.00513334758579731, + 0.008948124013841152, + 0.004495433531701565, + 0.00392268318682909, + 0.008513176813721657, + 0.001986695220693946, + 0.002147823804989457, + 0.0035496982745826244, + 0.004334202967584133, + 0.002503315918147564, + 0.001976415980607271, + 0.001925082178786397, + 0.0026569454930722713, + 0.00289433216676116, + 0.00966041348874569, + 0.002530118217691779, + 0.002622896572574973, + 0.0038942042738199234, + 0.010487827472388744, + 0.002182763535529375, + 0.0038885518442839384, + 0.02375439740717411, + 0.10185074806213379, + 0.01462673768401146, + 0.012888266704976559, + 0.016419660300016403, + 0.09023376554250717, + 0.0055805291049182415, + 0.012296547181904316, + 0.002521383808925748, + 0.005643099080771208, + 0.002531994367018342, + 0.0018607999663800001, + 0.005782757885754108, + 0.005910808686167002, + 0.0027172863483428955, + 0.0043445792980492115, + 0.008477224968373775, + 0.018777934834361076, + 0.009163913317024708, + 0.006779517978429794, + 0.010874158702790737, + 0.014022685587406158, + 0.00725598307326436, + 0.007486608345061541, + 0.006486769765615463, + 0.027430854737758636, + 0.004226521123200655, + 0.00925098080188036, + 0.0037814946845173836, + 0.02353387512266636, + 0.0022478026803582907, + 0.0027955244295299053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.005958176217973232 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190716668963432 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190716668963432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190716668963432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41186991333961487, + 0.04261131212115288, + 0.04397241026163101, + 0.0337412990629673, + 0.018621915951371193, + 0.02929636463522911, + 0.2589576542377472, + 0.017420129850506783, + 0.02729552984237671, + 0.11621353030204773 + ], + "credit_map": [ + 0.04104059562087059, + 0.040675610303878784, + 0.06788357347249985, + 0.09564658254384995, + 0.0482390932738781, + 0.03661807253956795, + 0.03500722721219063, + 0.046759117394685745, + 0.0044697183184325695, + 0.004507820587605238, + 0.005748014897108078, + 0.006709539331495762, + 0.005531300324946642, + 0.003784453496336937, + 0.0040315864607691765, + 0.007828875444829464, + 0.005606194958090782, + 0.004472787957638502, + 0.005435263272374868, + 0.005245395470410585, + 0.008971994742751122, + 0.0039162226021289825, + 0.0042998818680644035, + 0.006024664733558893, + 0.004375878721475601, + 0.0043494380079209805, + 0.00390625512227416, + 0.0036189223174005747, + 0.005857743788510561, + 0.004380954895168543, + 0.0028367547784000635, + 0.004415351897478104, + 0.0017819611821323633, + 0.002646442735567689, + 0.002645385917276144, + 0.0028895875439047813, + 0.002650944981724024, + 0.0021334723569452763, + 0.0017461777897551656, + 0.002127942629158497, + 0.00203129556030035, + 0.008332351222634315, + 0.001956179039552808, + 0.002753633540123701, + 0.0034811224322766066, + 0.007269935682415962, + 0.001692972844466567, + 0.0017788702389225364, + 0.019975503906607628, + 0.09976405650377274, + 0.011007130146026611, + 0.009317458607256413, + 0.014131858013570309, + 0.09104561805725098, + 0.0038944219704717398, + 0.009821570478379726, + 0.001915865927003324, + 0.0015395989175885916, + 0.0025472298730164766, + 0.0017832026351243258, + 0.0021494266111403704, + 0.002120256656780839, + 0.0012732738396152854, + 0.0040912749245762825, + 0.002777348505333066, + 0.004581429064273834, + 0.0031239099334925413, + 0.00406158110126853, + 0.003004799596965313, + 0.0036147397477179766, + 0.0030747116543352604, + 0.00305700721219182, + 0.010132381692528725, + 0.043200813233852386, + 0.0056209503673017025, + 0.010609054006636143, + 0.005290284752845764, + 0.034647293388843536, + 0.003344126744195819, + 0.0033686142414808273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.014190716668963432 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292701959609985 + ], + "opposite_logprob_tokens": [ + -2.237926959991455 + ], + "logprob_reward": -0.11292701959609985, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292701959609985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13141554594039917, + 0.05420491099357605, + 0.18387840688228607, + 0.06044093891978264, + 0.059603992849588394, + 0.0968795195221901, + 0.14050456881523132, + 0.10178928077220917, + 0.0778588056564331, + 0.09342404454946518 + ], + "credit_map": [ + 0.010699763894081116, + 0.013253225944936275, + 0.025473367422819138, + 0.03720890358090401, + 0.0148010840639472, + 0.008322054520249367, + 0.00952625460922718, + 0.012130888178944588, + 0.007416729815304279, + 0.0066985334269702435, + 0.005428400356322527, + 0.005332606844604015, + 0.010871803387999535, + 0.005300548858940601, + 0.006717503536492586, + 0.006438783369958401, + 0.031094668433070183, + 0.02550114504992962, + 0.023243103176355362, + 0.017518524080514908, + 0.024538343772292137, + 0.02263149619102478, + 0.016346680000424385, + 0.023004453629255295, + 0.005958189256489277, + 0.006689738482236862, + 0.0064384727738797665, + 0.00743942242115736, + 0.013511723838746548, + 0.008664854802191257, + 0.00628072302788496, + 0.005457817576825619, + 0.005516418721526861, + 0.008067278191447258, + 0.0060970354825258255, + 0.008972152136266232, + 0.013752289116382599, + 0.006389257963746786, + 0.005516849923878908, + 0.0052927155047655106, + 0.008756320923566818, + 0.023862725123763084, + 0.00740095553919673, + 0.008505193516612053, + 0.00844669621437788, + 0.021686606109142303, + 0.008125558495521545, + 0.010095464996993542, + 0.006860086228698492, + 0.04566699266433716, + 0.00799534097313881, + 0.005545788444578648, + 0.01342068426311016, + 0.05316454917192459, + 0.004151033703237772, + 0.0037000877782702446, + 0.00913811381906271, + 0.026535775512456894, + 0.004481689538806677, + 0.00510176457464695, + 0.015593531541526318, + 0.020730460062623024, + 0.012195223942399025, + 0.008012725040316582, + 0.005708574317395687, + 0.009356490336358547, + 0.006417101714760065, + 0.009243237785995007, + 0.01335964910686016, + 0.010417068377137184, + 0.01762584038078785, + 0.005730837117880583, + 0.007859352976083755, + 0.030406741425395012, + 0.006890385411679745, + 0.012354177422821522, + 0.005869871936738491, + 0.022641891613602638, + 0.0039057794492691755, + 0.0034958457108587027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.11292701959609985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot tilts steadily upward, its movement synchronized with the powerful rise of a single, massive blade from a wind turbine. Made of sleek gray-white fiberglass, the blade fills the frame, its smooth, aerodynamic surface capturing the first golden rays of the rising sun, forming a brilliant leading edge against the deep blue of the clear morning sky. As it reaches its peak, the camera continues to follow its relentless, hypnotic rotation, cutting cleanly through the air with powerful grace, its tip slightly blurred by sheer speed. Below, rolling green hills, vibrant and lush, stretch to the horizon, while a gentle morning mist gathers softly in the valley. The scene is bathed in the cool, clear light of dawn, creating long, soft shadows and evoking an atmosphere of majestic power and tranquil, clean energy.", + "seed": 20835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20835_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": -0.09773946367204189, + "vlm_logprob": -0.09773946367204189, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.09773946367204189, + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04042259231209755 + ], + "opposite_logprob_tokens": [ + -3.4154226779937744 + ], + "logprob_reward": -0.04042259231209755, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04042259231209755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08545568585395813, + 0.039537038654088974, + 0.03667660802602768, + 0.031968533992767334, + 0.0357719324529171, + 0.07068714499473572, + 0.06475178152322769, + 0.15449298918247223, + 0.2466895878314972, + 0.23396866023540497 + ], + "credit_map": [ + 0.012569166719913483, + 0.010931757278740406, + 0.01136342529207468, + 0.011844403110444546, + 0.00889267586171627, + 0.009356686845421791, + 0.010436397977173328, + 0.010061176493763924, + 0.004897222854197025, + 0.0031700925901532173, + 0.005385161377489567, + 0.004393340088427067, + 0.003493145341053605, + 0.007466444745659828, + 0.004989347420632839, + 0.005742281675338745, + 0.005071455612778664, + 0.003080887021496892, + 0.004399676341563463, + 0.005197760183364153, + 0.003946639597415924, + 0.004458865616470575, + 0.0037942330818623304, + 0.006727092433720827, + 0.005377295892685652, + 0.004180323798209429, + 0.0037097930908203125, + 0.004524354822933674, + 0.0035213432274758816, + 0.002654512645676732, + 0.004186432808637619, + 0.00381447677500546, + 0.0043468470685184, + 0.003330803941935301, + 0.005729572381824255, + 0.003288386156782508, + 0.0067230528220534325, + 0.004067928530275822, + 0.005484122317284346, + 0.0028012190014123917, + 0.005243036895990372, + 0.012635812163352966, + 0.011676161549985409, + 0.0076688919216394424, + 0.009655963629484177, + 0.009330911561846733, + 0.0077126543037593365, + 0.006763715762645006, + 0.006480826530605555, + 0.012053724378347397, + 0.00850057415664196, + 0.006611873861402273, + 0.007694396190345287, + 0.006887135095894337, + 0.008116049692034721, + 0.008407196961343288, + 0.014340263791382313, + 0.016460077837109566, + 0.02111203409731388, + 0.025344019755721092, + 0.024238847196102142, + 0.02676800638437271, + 0.009441699832677841, + 0.016788050532341003, + 0.025387320667505264, + 0.029467778280377388, + 0.016841091215610504, + 0.030726052820682526, + 0.057303912937641144, + 0.02973899431526661, + 0.020500684157013893, + 0.03672376275062561, + 0.017386827617883682, + 0.018786277621984482, + 0.06105871871113777, + 0.05273841321468353, + 0.022215461358428, + 0.017592739313840866, + 0.01662319153547287, + 0.027567002922296524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.04042259231209755 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704554200172424 + ], + "opposite_logprob_tokens": [ + -2.1270456314086914 + ], + "logprob_reward": -0.12704554200172424, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704554200172424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29792845249176025, + 0.04071146622300148, + 0.15029364824295044, + 0.05130010470747948, + 0.03516784682869911, + 0.06234356015920639, + 0.0991201400756836, + 0.04945550113916397, + 0.11463434994220734, + 0.09904486685991287 + ], + "credit_map": [ + 0.04046536982059479, + 0.042516354471445084, + 0.04218253493309021, + 0.05092565342783928, + 0.021279064938426018, + 0.01990140788257122, + 0.04645919054746628, + 0.03419890254735947, + 0.006746537517756224, + 0.0032853225711733103, + 0.00616058986634016, + 0.005004982464015484, + 0.0035254235845059156, + 0.00591621408239007, + 0.005571173038333654, + 0.004501227289438248, + 0.02337966300547123, + 0.011199341155588627, + 0.027093706652522087, + 0.0166667178273201, + 0.01403356809169054, + 0.021222703158855438, + 0.015099849551916122, + 0.021598104387521744, + 0.005660508293658495, + 0.004861319903284311, + 0.008219098672270775, + 0.006191522814333439, + 0.011057122610509396, + 0.003957820124924183, + 0.00695025222375989, + 0.004402463324368, + 0.005125861614942551, + 0.004184728022664785, + 0.006270228419452906, + 0.004108532331883907, + 0.0055526322685182095, + 0.0026145433075726032, + 0.004899130668491125, + 0.00241219368763268, + 0.008391528390347958, + 0.0087943309918046, + 0.010522349737584591, + 0.009109403006732464, + 0.00713927298784256, + 0.005439808126538992, + 0.007458087056875229, + 0.005488785915076733, + 0.006945509929209948, + 0.018466811627149582, + 0.012960626743733883, + 0.010107648558914661, + 0.0076816449873149395, + 0.010789972729980946, + 0.01769614964723587, + 0.014471788890659809, + 0.0039208983071148396, + 0.005750028882175684, + 0.006397558376193047, + 0.008451947011053562, + 0.0074439519084990025, + 0.0051031699404120445, + 0.004692834801971912, + 0.007695114705711603, + 0.010369821451604366, + 0.012747476808726788, + 0.010578664019703865, + 0.018495403230190277, + 0.011440680362284184, + 0.013674246147274971, + 0.012444846332073212, + 0.024883227422833443, + 0.009350702166557312, + 0.00751496572047472, + 0.02311853878200054, + 0.0220255758613348, + 0.007986939512193203, + 0.009592082351446152, + 0.00831926055252552, + 0.011136811226606369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.12704554200172424 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748029962182045 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748029962182045, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748029962182045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586177498102188, + 0.10440307855606079, + 0.08486859500408173, + 0.026320679113268852, + 0.05227901041507721, + 0.05447366461157799, + 0.04639323428273201, + 0.07093332707881927, + 0.28072988986968994, + 0.12098078429698944 + ], + "credit_map": [ + 0.021006213501095772, + 0.021589985117316246, + 0.020838990807533264, + 0.022443464025855064, + 0.010888049378991127, + 0.014177443459630013, + 0.029266182333230972, + 0.01840742863714695, + 0.012538637034595013, + 0.008571352809667587, + 0.012337886728346348, + 0.009974421933293343, + 0.008566764183342457, + 0.019497031345963478, + 0.012117788195610046, + 0.020799199119210243, + 0.014238085597753525, + 0.006654669065028429, + 0.014322018250823021, + 0.009479692205786705, + 0.007482683751732111, + 0.010894141159951687, + 0.009011977352201939, + 0.01278532762080431, + 0.002315361285582185, + 0.002473351778462529, + 0.004347654990851879, + 0.004241183865815401, + 0.0034741880372166634, + 0.0031017360743135214, + 0.003463828470557928, + 0.002903374843299389, + 0.005814061034470797, + 0.006540680769830942, + 0.009228277951478958, + 0.0033784527331590652, + 0.010407749563455582, + 0.005959658417850733, + 0.006390540394932032, + 0.004559593740850687, + 0.004124798811972141, + 0.008899234235286713, + 0.00928326603025198, + 0.0048230295069515705, + 0.007610707078129053, + 0.0063168699853122234, + 0.006308836862444878, + 0.007106923498213291, + 0.003813626477494836, + 0.009410257451236248, + 0.005410796497017145, + 0.007429553661495447, + 0.004700865130871534, + 0.004284439142793417, + 0.005448667332530022, + 0.005895028356462717, + 0.006948125082999468, + 0.009962007403373718, + 0.009325508028268814, + 0.009269705973565578, + 0.012513702735304832, + 0.01002086978405714, + 0.0055345939472317696, + 0.007358811795711517, + 0.027583805844187737, + 0.0346078984439373, + 0.021834485232830048, + 0.04533318430185318, + 0.05129317194223404, + 0.025896495208144188, + 0.026358654722571373, + 0.04782221093773842, + 0.009220805019140244, + 0.012064329348504543, + 0.02032141573727131, + 0.016592783853411674, + 0.014834240078926086, + 0.012257230468094349, + 0.009823717176914215, + 0.025866268202662468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.006748029962182045 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293703317642212 + ], + "opposite_logprob_tokens": [ + -2.2379369735717773 + ], + "logprob_reward": -0.11293703317642212, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293703317642212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034834958612918854, + 0.02322573959827423, + 0.04766589775681496, + 0.014991065487265587, + 0.019741222262382507, + 0.020766019821166992, + 0.019205408170819283, + 0.09546180814504623, + 0.48745197057724, + 0.23665587604045868 + ], + "credit_map": [ + 0.0048543126322329044, + 0.004343727603554726, + 0.004634384997189045, + 0.005524998065084219, + 0.0028801830485463142, + 0.002590401330962777, + 0.006051955278962851, + 0.003954999148845673, + 0.0038267080672085285, + 0.002045808359980583, + 0.0029753237031400204, + 0.0025119916535913944, + 0.0016702164430171251, + 0.0032798219472169876, + 0.002545927884057164, + 0.004369943868368864, + 0.008296768181025982, + 0.0035881472285836935, + 0.008825554512441158, + 0.005211941432207823, + 0.004187903366982937, + 0.006170481443405151, + 0.004944200161844492, + 0.006440904922783375, + 0.0015504828188568354, + 0.001736894017085433, + 0.0025183514226228, + 0.0019107313128188252, + 0.001657150569371879, + 0.001587319653481245, + 0.002386385342106223, + 0.0016437520971521735, + 0.00277309468947351, + 0.002187848323956132, + 0.003688220866024494, + 0.0016270304331555963, + 0.004017379134893417, + 0.002006496535614133, + 0.0018692314624786377, + 0.0015719217481091619, + 0.002211445476859808, + 0.0020886182319372892, + 0.0034221874084323645, + 0.0024612178094685078, + 0.002794982399791479, + 0.002649392234161496, + 0.0025385823100805283, + 0.002599595580250025, + 0.0023833061568439007, + 0.003820116864517331, + 0.002657488686963916, + 0.0019344388274475932, + 0.002153236186131835, + 0.0023236016277223825, + 0.0022177877835929394, + 0.0017154334345832467, + 0.0073954383842647076, + 0.008223365992307663, + 0.024138862267136574, + 0.010279355570673943, + 0.009651781991124153, + 0.021146465092897415, + 0.005346993450075388, + 0.009279548190534115, + 0.048616375774145126, + 0.06295088678598404, + 0.03833027184009552, + 0.06543496996164322, + 0.09909293800592422, + 0.049590568989515305, + 0.049259964376688004, + 0.07417603582143784, + 0.01743660494685173, + 0.027003051713109016, + 0.04408537223935127, + 0.02976791001856327, + 0.02448302134871483, + 0.023987475782632828, + 0.0192856602370739, + 0.05060679093003273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.11293703317642212 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015441209077835 + ], + "opposite_logprob_tokens": [ + -1.701544165611267 + ], + "logprob_reward": -0.2015441209077835, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015441209077835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08332125097513199, + 0.03083636425435543, + 0.13512447476387024, + 0.05590769276022911, + 0.037558168172836304, + 0.062103111296892166, + 0.1585240364074707, + 0.06744706630706787, + 0.1987803429365158, + 0.17039749026298523 + ], + "credit_map": [ + 0.012647108174860477, + 0.013383391313254833, + 0.012527654878795147, + 0.010774929076433182, + 0.0075589744374156, + 0.005751719232648611, + 0.009004044346511364, + 0.011673424392938614, + 0.00415361113846302, + 0.00291959079913795, + 0.004077652469277382, + 0.004422105848789215, + 0.002925467910245061, + 0.005013606511056423, + 0.0028460954781621695, + 0.004478237126022577, + 0.019370296970009804, + 0.015653623268008232, + 0.019610300660133362, + 0.014327323995530605, + 0.011374284513294697, + 0.015681372955441475, + 0.019212789833545685, + 0.01989448256790638, + 0.006607505958527327, + 0.006920584011822939, + 0.008595532737672329, + 0.010036250576376915, + 0.007981500588357449, + 0.0048310477286577225, + 0.005931565538048744, + 0.005003707017749548, + 0.007186558563262224, + 0.003938279114663601, + 0.0077378638088703156, + 0.004080554004758596, + 0.005832954775542021, + 0.003365582786500454, + 0.0031321323476731777, + 0.002284242305904627, + 0.005865585990250111, + 0.009522415697574615, + 0.011827727779746056, + 0.007430912461131811, + 0.010241547599434853, + 0.0058938078582286835, + 0.0067947558127343655, + 0.004526354372501373, + 0.016740920022130013, + 0.03562378138303757, + 0.0159921832382679, + 0.021253397688269615, + 0.01687552221119404, + 0.013384847901761532, + 0.02119462750852108, + 0.017458755522966385, + 0.0059908186085522175, + 0.008823028765618801, + 0.00915104616433382, + 0.009123899042606354, + 0.01038700807839632, + 0.008184143342077732, + 0.006430323701351881, + 0.009356801398098469, + 0.01876913756132126, + 0.029494598507881165, + 0.014378389343619347, + 0.021773062646389008, + 0.04940731078386307, + 0.028517162427306175, + 0.014627879485487938, + 0.021812794730067253, + 0.01295713521540165, + 0.019638143479824066, + 0.031112870201468468, + 0.02828315645456314, + 0.016360241919755936, + 0.015751587226986885, + 0.014041777700185776, + 0.03225257620215416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.2015441209077835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot tilts steadily upward, its movement synchronized with the powerful rise of a single, massive blade from a wind turbine. Made of sleek gray-white fiberglass, the blade fills the frame, its smooth, aerodynamic surface capturing the first golden rays of the rising sun, forming a brilliant leading edge against the deep blue of the clear morning sky. As it reaches its peak, the camera continues to follow its relentless, hypnotic rotation, cutting cleanly through the air with powerful grace, its tip slightly blurred by sheer speed. Below, rolling green hills, vibrant and lush, stretch to the horizon, while a gentle morning mist gathers softly in the valley. The scene is bathed in the cool, clear light of dawn, creating long, soft shadows and evoking an atmosphere of majestic power and tranquil, clean energy.", + "seed": 20836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20836_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": -0.0211868932005018, + "vlm_logprob": -0.0211868932005018, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.0211868932005018, + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027315936982631683 + ], + "opposite_logprob_tokens": [ + -3.902315855026245 + ], + "logprob_reward": -0.027315936982631683, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027315936982631683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12906906008720398, + 0.0759979784488678, + 0.09056898951530457, + 0.09845364093780518, + 0.2552337348461151, + 0.1507718414068222, + 0.06443016231060028, + 0.02162894420325756, + 0.0380178727209568, + 0.07582782208919525 + ], + "credit_map": [ + 0.017190011218190193, + 0.020482072606682777, + 0.012129721231758595, + 0.01877778209745884, + 0.016541417688131332, + 0.014517633244395256, + 0.014764797873795033, + 0.01466561108827591, + 0.010451092384755611, + 0.009392942301928997, + 0.01170725654810667, + 0.009503780864179134, + 0.007254826836287975, + 0.010605255141854286, + 0.006602919194847345, + 0.010479903779923916, + 0.013193413615226746, + 0.012312717735767365, + 0.013635356910526752, + 0.006941043771803379, + 0.01107105053961277, + 0.009958496317267418, + 0.009769083000719547, + 0.013687831349670887, + 0.006875135470181704, + 0.009530610404908657, + 0.03207921236753464, + 0.026800772175192833, + 0.008131834678351879, + 0.005555504932999611, + 0.0042210849933326244, + 0.005259478464722633, + 0.050472892820835114, + 0.034434158354997635, + 0.053345173597335815, + 0.04397502914071083, + 0.015451380051672459, + 0.015321636572480202, + 0.025947634130716324, + 0.016285860911011696, + 0.013207192532718182, + 0.016231996938586235, + 0.022348014637827873, + 0.027726856991648674, + 0.015182248316705227, + 0.017217546701431274, + 0.02198825776576996, + 0.016869736835360527, + 0.005775795318186283, + 0.00802555400878191, + 0.009620187804102898, + 0.013578040525317192, + 0.004301805980503559, + 0.0036073888186365366, + 0.007995042949914932, + 0.01152634434401989, + 0.0024901018477976322, + 0.0027358513325452805, + 0.002251038560643792, + 0.004968185909092426, + 0.0019935164600610733, + 0.0013621750986203551, + 0.0018878878327086568, + 0.003940188325941563, + 0.0045275334268808365, + 0.008009867742657661, + 0.00377286272123456, + 0.0057522947899997234, + 0.004539647605270147, + 0.0028122246731072664, + 0.0040973760187625885, + 0.00450606644153595, + 0.006066595204174519, + 0.02720741368830204, + 0.011751196347177029, + 0.006599684711545706, + 0.015427743084728718, + 0.0035002236254513264, + 0.0021209612023085356, + 0.003154006088152528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.027315936982631683 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081814765930176 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081814765930176, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081814765930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07202565670013428, + 0.07143603265285492, + 0.028896620497107506, + 0.10783771425485611, + 0.3223782777786255, + 0.10208035260438919, + 0.09196598827838898, + 0.06375175714492798, + 0.09495089203119278, + 0.044676683843135834 + ], + "credit_map": [ + 0.005658864509314299, + 0.004566575866192579, + 0.005859020631760359, + 0.007466654758900404, + 0.00992080569267273, + 0.015510141849517822, + 0.013547349721193314, + 0.009496246464550495, + 0.009529623202979565, + 0.007466297131031752, + 0.009361411444842815, + 0.0074087996035814285, + 0.006197952665388584, + 0.014389201998710632, + 0.008641809225082397, + 0.008440935052931309, + 0.006307181436568499, + 0.00299716810695827, + 0.004884224850684404, + 0.0028067617677152157, + 0.002237791894003749, + 0.0029634875245392323, + 0.002941625192761421, + 0.0037583804223686457, + 0.012792280875146389, + 0.02663601189851761, + 0.02786889486014843, + 0.02188423089683056, + 0.005734873469918966, + 0.005070456303656101, + 0.004383052699267864, + 0.0034679160453379154, + 0.07822080701589584, + 0.05063827708363533, + 0.08359377831220627, + 0.06464338302612305, + 0.011487273499369621, + 0.008657488040626049, + 0.017413359135389328, + 0.00772391352802515, + 0.012188092805445194, + 0.0162671972066164, + 0.009652567096054554, + 0.014106585644185543, + 0.012872696854174137, + 0.01251633558422327, + 0.013934447430074215, + 0.010542429983615875, + 0.014052700251340866, + 0.026311349123716354, + 0.009351431392133236, + 0.008503665216267109, + 0.013219466432929039, + 0.006462625227868557, + 0.0061536431312561035, + 0.007911105640232563, + 0.007352111395448446, + 0.013527394272387028, + 0.005202592350542545, + 0.010804744437336922, + 0.007619050797075033, + 0.00466670049354434, + 0.004326066002249718, + 0.010253094136714935, + 0.010316019877791405, + 0.022071050480008125, + 0.009639528580009937, + 0.012226897291839123, + 0.013051965273916721, + 0.006966965738683939, + 0.009800340048968792, + 0.010878130793571472, + 0.007437142077833414, + 0.00914053339511156, + 0.004985950421541929, + 0.005719540640711784, + 0.006817838177084923, + 0.0021051540970802307, + 0.003018693532794714, + 0.005451834294945002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.011081814765930176 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823891816660762 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002823891816660762, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823891816660762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20733480155467987, + 0.14320997893810272, + 0.06803160905838013, + 0.13252609968185425, + 0.13431942462921143, + 0.07614284008741379, + 0.06854821741580963, + 0.053474247455596924, + 0.06562742590904236, + 0.0507853664457798 + ], + "credit_map": [ + 0.018209204077720642, + 0.015995245426893234, + 0.018020091578364372, + 0.02196924388408661, + 0.022553838789463043, + 0.04409201443195343, + 0.03927998244762421, + 0.02721518650650978, + 0.014535618014633656, + 0.00923595204949379, + 0.03449459373950958, + 0.012267849408090115, + 0.008402315899729729, + 0.022443844005465508, + 0.012440205551683903, + 0.029389601200819016, + 0.010148102417588234, + 0.010035793296992779, + 0.01129577774554491, + 0.006831918377429247, + 0.006362622603774071, + 0.006266618613153696, + 0.007460517343133688, + 0.009630253538489342, + 0.0184251107275486, + 0.017933165654540062, + 0.04236843064427376, + 0.0388462170958519, + 0.0034029833041131496, + 0.003784729400649667, + 0.0036535451654344797, + 0.004111906047910452, + 0.023525768890976906, + 0.01624678075313568, + 0.028186537325382233, + 0.02562897466123104, + 0.00627962127327919, + 0.010669604875147343, + 0.013774912804365158, + 0.010007231496274471, + 0.008939847350120544, + 0.009490925818681717, + 0.012200990691781044, + 0.012364761903882027, + 0.007904179394245148, + 0.00852407980710268, + 0.009190192446112633, + 0.007527862209826708, + 0.008983776904642582, + 0.01572735421359539, + 0.0054527511820197105, + 0.008469636552035809, + 0.008190704509615898, + 0.004826115909963846, + 0.005409277975559235, + 0.01148859690874815, + 0.004205420613288879, + 0.009172958321869373, + 0.0037772278301417828, + 0.006886376533657312, + 0.013002300634980202, + 0.00558072654530406, + 0.0037897760048508644, + 0.007059459108859301, + 0.0075932578183710575, + 0.014279638417065144, + 0.005863073747605085, + 0.00878484733402729, + 0.010369791649281979, + 0.00320479366928339, + 0.00795335229486227, + 0.007578665856271982, + 0.0049797603860497475, + 0.018167467787861824, + 0.004396155942231417, + 0.003274941584095359, + 0.011683897115290165, + 0.0024576468858867884, + 0.002776402048766613, + 0.0030490956269204617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002823891816660762 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779401123523712 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779401123523712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779401123523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11063503473997116, + 0.15729181468486786, + 0.055137280374765396, + 0.24015040695667267, + 0.10640585422515869, + 0.1181369349360466, + 0.06687449663877487, + 0.05713511258363724, + 0.04533668980002403, + 0.04289639741182327 + ], + "credit_map": [ + 0.01238393783569336, + 0.010440933518111706, + 0.009834770113229752, + 0.014043875969946384, + 0.010466311126947403, + 0.024592988193035126, + 0.014594607055187225, + 0.014277609996497631, + 0.019365152344107628, + 0.011977825313806534, + 0.027068115770816803, + 0.011724617332220078, + 0.012823151424527168, + 0.027262847870588303, + 0.019952667877078056, + 0.02711743488907814, + 0.00806346070021391, + 0.00873289629817009, + 0.0086243050172925, + 0.0053399670869112015, + 0.004557599313557148, + 0.004099788144230843, + 0.006040362175554037, + 0.00967890303581953, + 0.03424343094229698, + 0.0623740591108799, + 0.059915754944086075, + 0.047725826501846313, + 0.010513164103031158, + 0.009683644399046898, + 0.0077695706859230995, + 0.00792495347559452, + 0.014503887854516506, + 0.010522466152906418, + 0.029247568920254707, + 0.02668176405131817, + 0.005897830706089735, + 0.0034671195317059755, + 0.011625668033957481, + 0.004459548741579056, + 0.010951738804578781, + 0.014689448289573193, + 0.019696472212672234, + 0.022582966834306717, + 0.010779423639178276, + 0.01207730919122696, + 0.01599567010998726, + 0.011363910511136055, + 0.007418330293148756, + 0.008246473968029022, + 0.006965266074985266, + 0.009536989033222198, + 0.006973910145461559, + 0.005368443671613932, + 0.008923457935452461, + 0.013441625982522964, + 0.006293091922998428, + 0.010789106599986553, + 0.0047278134152293205, + 0.0120565015822649, + 0.00613191444426775, + 0.003689137287437916, + 0.00470022764056921, + 0.00874731969088316, + 0.00570298358798027, + 0.007488475646823645, + 0.004510683473199606, + 0.006317117251455784, + 0.0065471818670630455, + 0.0026594316586852074, + 0.006149123888462782, + 0.005961693357676268, + 0.004920436069369316, + 0.012553154490888119, + 0.002968358574435115, + 0.00324481469579041, + 0.010214552283287048, + 0.0026903555262833834, + 0.002265178132802248, + 0.004039550665766001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.009779401123523712 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054933421313762665 + ], + "opposite_logprob_tokens": [ + -2.9299333095550537 + ], + "logprob_reward": -0.054933421313762665, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054933421313762665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09540696442127228, + 0.06669748574495316, + 0.050471480935811996, + 0.07964084297418594, + 0.12217877805233002, + 0.3443462550640106, + 0.07071688771247864, + 0.07697711139917374, + 0.06098183989524841, + 0.03258228302001953 + ], + "credit_map": [ + 0.008287551812827587, + 0.00787535309791565, + 0.00807681679725647, + 0.010841031558811665, + 0.012163345701992512, + 0.019424451515078545, + 0.01766946166753769, + 0.011068958789110184, + 0.00949622131884098, + 0.008020210079848766, + 0.010090903379023075, + 0.008533112704753876, + 0.005914292763918638, + 0.009056351147592068, + 0.006327333860099316, + 0.009259060025215149, + 0.007951455190777779, + 0.006120176054537296, + 0.006747722160071135, + 0.004065610934048891, + 0.005863908678293228, + 0.006701342295855284, + 0.00507713295519352, + 0.007944133132696152, + 0.008200819604098797, + 0.009708055295050144, + 0.02239600569009781, + 0.02073679119348526, + 0.005743717774748802, + 0.004435639828443527, + 0.0035202056169509888, + 0.004899602849036455, + 0.02819014899432659, + 0.018125582486391068, + 0.02861860953271389, + 0.024463670328259468, + 0.0038770108949393034, + 0.00451516080647707, + 0.008973658084869385, + 0.005414940416812897, + 0.042020201683044434, + 0.05429242551326752, + 0.05515127256512642, + 0.06415670365095139, + 0.0297753494232893, + 0.022129910066723824, + 0.04277963936328888, + 0.03404078632593155, + 0.006917821709066629, + 0.00823147315531969, + 0.00965911615639925, + 0.0135815953835845, + 0.005715962033718824, + 0.004826721735298634, + 0.007626688107848167, + 0.014157513156533241, + 0.008733704686164856, + 0.015302835032343864, + 0.006624873727560043, + 0.01421419344842434, + 0.010229073464870453, + 0.0038592976052314043, + 0.005912321154028177, + 0.012100813910365105, + 0.006747442297637463, + 0.016370799392461777, + 0.005152062512934208, + 0.007387596648186445, + 0.010747651569545269, + 0.003220711601898074, + 0.005229108966886997, + 0.00612646946683526, + 0.004184464458376169, + 0.009392784908413887, + 0.0030451049096882343, + 0.0025331112556159496, + 0.006658113561570644, + 0.0016631287289783359, + 0.0021505397744476795, + 0.0029550353065133095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.054933421313762665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle shot tilts steadily upward, its movement synchronized with the powerful rise of a single, massive blade from a wind turbine. Made of sleek gray-white fiberglass, the blade fills the frame, its smooth, aerodynamic surface capturing the first golden rays of the rising sun, forming a brilliant leading edge against the deep blue of the clear morning sky. As it reaches its peak, the camera continues to follow its relentless, hypnotic rotation, cutting cleanly through the air with powerful grace, its tip slightly blurred by sheer speed. Below, rolling green hills, vibrant and lush, stretch to the horizon, while a gentle morning mist gathers softly in the valley. The scene is bathed in the cool, clear light of dawn, creating long, soft shadows and evoking an atmosphere of majestic power and tranquil, clean energy.", + "seed": 20837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_10/index_126_seed_20837_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": -0.03162076650187373, + "vlm_logprob": -0.03162076650187373, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03162076650187373, + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026533694937825203 + ], + "opposite_logprob_tokens": [ + -3.901533603668213 + ], + "logprob_reward": -0.026533694937825203, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026533694937825203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3700929582118988, + 0.07242021709680557, + 0.08694690465927124, + 0.030132973566651344, + 0.10745556652545929, + 0.07045382261276245, + 0.06470541656017303, + 0.07952350378036499, + 0.036727599799633026, + 0.081541046500206 + ], + "credit_map": [ + 0.030201662331819534, + 0.036717236042022705, + 0.056878186762332916, + 0.061011891812086105, + 0.039535488933324814, + 0.05016137659549713, + 0.052571896463632584, + 0.04301521182060242, + 0.00811753049492836, + 0.008375082165002823, + 0.010212257504463196, + 0.010793078690767288, + 0.009188485331833363, + 0.00856095738708973, + 0.007883884012699127, + 0.009288939647376537, + 0.008716655895113945, + 0.010809476487338543, + 0.009391529485583305, + 0.010856358334422112, + 0.016119446605443954, + 0.010632717981934547, + 0.010619710199534893, + 0.009800995700061321, + 0.0031471720430999994, + 0.004267903510481119, + 0.004106235224753618, + 0.0035883388482034206, + 0.003968091681599617, + 0.004892704542726278, + 0.002792962361127138, + 0.003369562327861786, + 0.010185180231928825, + 0.014582066796720028, + 0.01535490807145834, + 0.008663809858262539, + 0.016604101285338402, + 0.01256908755749464, + 0.00906197540462017, + 0.0204344242811203, + 0.00907902605831623, + 0.01120244711637497, + 0.009129418060183525, + 0.005952162202447653, + 0.017861666157841682, + 0.006239859852939844, + 0.00510809849947691, + 0.005881139077246189, + 0.00744358915835619, + 0.00749741168692708, + 0.010083012282848358, + 0.005241776816546917, + 0.008449510671198368, + 0.007410716265439987, + 0.007308777421712875, + 0.01127061527222395, + 0.005608593113720417, + 0.016003184020519257, + 0.016611702740192413, + 0.010187320411205292, + 0.008175382390618324, + 0.007102493196725845, + 0.006877013947814703, + 0.008957816287875175, + 0.0027933029923588037, + 0.004457533825188875, + 0.005121795926243067, + 0.006941828411072493, + 0.005850190296769142, + 0.002961951307952404, + 0.00443509453907609, + 0.004165900871157646, + 0.006392249371856451, + 0.012793918140232563, + 0.007397171575576067, + 0.010979521088302135, + 0.018354913219809532, + 0.008323248475790024, + 0.009955578483641148, + 0.007344442885369062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.026533694937825203 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003626082558184862 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.003626082558184862, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003626082558184862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24766714870929718, + 0.04204409569501877, + 0.11676271259784698, + 0.06098513677716255, + 0.11371957510709763, + 0.03533357009291649, + 0.043556053191423416, + 0.11784031242132187, + 0.19702893495559692, + 0.02506254054605961 + ], + "credit_map": [ + 0.02752428501844406, + 0.05621996149420738, + 0.030344456434249878, + 0.04374208301305771, + 0.019139861688017845, + 0.02890516072511673, + 0.020285554230213165, + 0.02150573395192623, + 0.004157263785600662, + 0.005743259098380804, + 0.009610201232135296, + 0.005872262641787529, + 0.0031376697588711977, + 0.005859074182808399, + 0.00294978404417634, + 0.004714576993137598, + 0.012082578614354134, + 0.01529509574174881, + 0.014586806297302246, + 0.016434134915471077, + 0.014753100462257862, + 0.02027926780283451, + 0.011487839743494987, + 0.011843876913189888, + 0.005911021959036589, + 0.007757240440696478, + 0.009991704486310482, + 0.007196773774921894, + 0.009218638762831688, + 0.007909337058663368, + 0.005327430088073015, + 0.007672984153032303, + 0.008786933496594429, + 0.01358796376734972, + 0.01424640417098999, + 0.010406723245978355, + 0.014330912381410599, + 0.01479580719023943, + 0.0093763442710042, + 0.02818848192691803, + 0.003961192909628153, + 0.004302503541111946, + 0.00961229670792818, + 0.005072569940239191, + 0.004051183816045523, + 0.002873410237953067, + 0.00203310395590961, + 0.00342730525881052, + 0.00511722220107913, + 0.00498998025432229, + 0.0073433853685855865, + 0.003692199010401964, + 0.004853839520365, + 0.005980417598038912, + 0.004908241331577301, + 0.006670764181762934, + 0.013812199234962463, + 0.017020056024193764, + 0.0187677089124918, + 0.01403498649597168, + 0.01709955371916294, + 0.010447102598845959, + 0.010309582576155663, + 0.01634911447763443, + 0.01674751564860344, + 0.03458191454410553, + 0.02567378617823124, + 0.03995858132839203, + 0.023122884333133698, + 0.014558093622326851, + 0.025505587458610535, + 0.01688055880367756, + 0.0026435726322233677, + 0.0028857060242444277, + 0.00389788462780416, + 0.002892749384045601, + 0.0033228425309062004, + 0.0029513719491660595, + 0.0024021496064960957, + 0.004066260997205973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.003626082558184862 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257708020508289 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257708020508289, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257708020508289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3359938859939575, + 0.083180733025074, + 0.12220660597085953, + 0.06230686977505684, + 0.08200109750032425, + 0.06206412613391876, + 0.030481697991490364, + 0.09414756298065186, + 0.06969424337148666, + 0.057923149317502975 + ], + "credit_map": [ + 0.023774946108460426, + 0.04180488362908363, + 0.09811393171548843, + 0.05424099043011665, + 0.022951485589146614, + 0.027426645159721375, + 0.03311848267912865, + 0.034562524408102036, + 0.008793269284069538, + 0.008479909040033817, + 0.014183995313942432, + 0.014088130556046963, + 0.009961026720702648, + 0.01052084006369114, + 0.0076137203723192215, + 0.009539843536913395, + 0.013062269426882267, + 0.01830274425446987, + 0.013174145482480526, + 0.013381344266235828, + 0.020760606974363327, + 0.013875393196940422, + 0.01582479290664196, + 0.013825309462845325, + 0.00595236849039793, + 0.008358282968401909, + 0.008960134349763393, + 0.006838540080934763, + 0.01046829391270876, + 0.00918559543788433, + 0.005616949871182442, + 0.006926705129444599, + 0.0067479875870049, + 0.009050967171788216, + 0.01114166621118784, + 0.008367924019694328, + 0.012152941897511482, + 0.008536247536540031, + 0.007761859335005283, + 0.018241506069898605, + 0.004453837405890226, + 0.006877283100038767, + 0.01475805975496769, + 0.008512812666594982, + 0.0064019314013421535, + 0.006087140645831823, + 0.005337691865861416, + 0.00963536649942398, + 0.0030564256012439728, + 0.004082437139004469, + 0.0044511170126497746, + 0.002347141969949007, + 0.0032947068102657795, + 0.0035503071267157793, + 0.0027004466392099857, + 0.006999116390943527, + 0.007200621534138918, + 0.02027442678809166, + 0.016911931335926056, + 0.010974954813718796, + 0.010206867940723896, + 0.008528065867722034, + 0.007934301160275936, + 0.0121163884177804, + 0.0070455134846270084, + 0.012488036416471004, + 0.008306358940899372, + 0.010860227979719639, + 0.008691515773534775, + 0.005954544525593519, + 0.009817592799663544, + 0.006530450191348791, + 0.00454273447394371, + 0.009030558168888092, + 0.006687648594379425, + 0.00874380860477686, + 0.011371933855116367, + 0.005300313234329224, + 0.006808809470385313, + 0.00543734198436141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.005257708020508289 + }, + { + "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 single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2786671817302704, + 0.09374016523361206, + 0.21871621906757355, + 0.08757511526346207, + 0.06849080324172974, + 0.06418127566576004, + 0.05664617195725441, + 0.06396651268005371, + 0.0487515963613987, + 0.01926497556269169 + ], + "credit_map": [ + 0.02300475724041462, + 0.022063620388507843, + 0.07192593812942505, + 0.0517766959965229, + 0.01865667663514614, + 0.03549443557858467, + 0.02459399215877056, + 0.031151067465543747, + 0.010053162463009357, + 0.01269436813890934, + 0.018960202112793922, + 0.016046803444623947, + 0.005517375655472279, + 0.012958040460944176, + 0.00722333462908864, + 0.01028687134385109, + 0.02459639310836792, + 0.028097407892346382, + 0.022687654942274094, + 0.028427312150597572, + 0.039085753262043, + 0.02334458939731121, + 0.027255402877926826, + 0.025221703574061394, + 0.006285786163061857, + 0.009606434032320976, + 0.02443278580904007, + 0.015424443408846855, + 0.004080161452293396, + 0.011684038676321507, + 0.003346350509673357, + 0.012715118005871773, + 0.005631077568978071, + 0.009174693375825882, + 0.013458353467285633, + 0.008390022441744804, + 0.006760549731552601, + 0.01142180897295475, + 0.0037518462631851435, + 0.009902456775307655, + 0.006574932485818863, + 0.007318783551454544, + 0.014323239214718342, + 0.008840429596602917, + 0.004069564864039421, + 0.007143269758671522, + 0.003494846634566784, + 0.012416210025548935, + 0.005699617322534323, + 0.007224926725029945, + 0.0076039559207856655, + 0.005547399166971445, + 0.0067782034166157246, + 0.006628494709730148, + 0.00622106809169054, + 0.01094250287860632, + 0.006367858964949846, + 0.010343615896999836, + 0.007458229549229145, + 0.007799305487424135, + 0.009989692829549313, + 0.005671295337378979, + 0.0064427731558680534, + 0.009893737733364105, + 0.003736581187695265, + 0.006849152967333794, + 0.007200703956186771, + 0.009849722497165203, + 0.007512630429118872, + 0.00393700459972024, + 0.00517831789329648, + 0.004487480502575636, + 0.001671023084782064, + 0.0035575723741203547, + 0.002015940845012665, + 0.0024886000901460648, + 0.003033241257071495, + 0.002068790839985013, + 0.002364757005125284, + 0.0020650497172027826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "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 main subject a single massive wind turbine blade made of sleek gray-white fiberglass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a low-angle shot and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade illuminated by golden rays of a rising sun, creating a brilliant leading edge against a deep blue clear sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are rolling green hills visible below the blade stretching to the horizon with gentle morning mist in the valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the blade tip appear slightly blurred due to its rapid rotation and speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129075363278389 + ], + "opposite_logprob_tokens": [ + -2.237907648086548 + ], + "logprob_reward": -0.1129075363278389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129075363278389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30140212178230286, + 0.036186911165714264, + 0.16598942875862122, + 0.09237132966518402, + 0.027211161330342293, + 0.02480420656502247, + 0.05764422193169594, + 0.2015620470046997, + 0.04854472354054451, + 0.04428388178348541 + ], + "credit_map": [ + 0.01930932141840458, + 0.018496688455343246, + 0.0746527835726738, + 0.057514581829309464, + 0.013395978137850761, + 0.04756857454776764, + 0.030493060126900673, + 0.039971113204956055, + 0.004296329338103533, + 0.005211613606661558, + 0.005836713593453169, + 0.0050175990909338, + 0.003436030587181449, + 0.0047541833482682705, + 0.0034310142509639263, + 0.004203425720334053, + 0.018525274470448494, + 0.02201075106859207, + 0.016983339563012123, + 0.022206099703907967, + 0.02775183506309986, + 0.019676877185702324, + 0.019087431952357292, + 0.019747823476791382, + 0.006968541070818901, + 0.009678815491497517, + 0.025715576484799385, + 0.014104390516877174, + 0.009200170636177063, + 0.010447625070810318, + 0.004659823141992092, + 0.011596388183534145, + 0.0022939720656722784, + 0.0031669072341173887, + 0.00419759564101696, + 0.0033154285047203302, + 0.002651298651471734, + 0.0039195045828819275, + 0.0022534749004989862, + 0.005412979982793331, + 0.0029618553817272186, + 0.003837164491415024, + 0.004645342472940683, + 0.002528389450162649, + 0.0015505643095821142, + 0.003538349410519004, + 0.0017229318618774414, + 0.004019607789814472, + 0.006414064671844244, + 0.007238243240863085, + 0.008906683884561062, + 0.0050370655953884125, + 0.00731356767937541, + 0.006531962193548679, + 0.005900336895138025, + 0.010302296839654446, + 0.01766107603907585, + 0.05929598584771156, + 0.04434245079755783, + 0.02295372076332569, + 0.010684231296181679, + 0.022233180701732635, + 0.009954896755516529, + 0.014436502940952778, + 0.005109165329486132, + 0.008795741945505142, + 0.005268947687000036, + 0.007474956568330526, + 0.005827695596963167, + 0.004161423537880182, + 0.00639351038262248, + 0.005513280164450407, + 0.004441743716597557, + 0.005556487012654543, + 0.0045911273919045925, + 0.006370831746608019, + 0.008339585736393929, + 0.0037572390865534544, + 0.007223601918667555, + 0.004003265872597694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.1129075363278389 + } + ], + "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/0000060/group_11/index_tensor([4205, 4205])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_tensor([4205, 4205])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..220592a13618208749fd4216d9ee9d106a1a2820 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_tensor([4205, 4205])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot frames the face of a young man, Moriya Ishida, caught in the center of a raging thunderstorm. The camera remains completely still, creating an intimate and unwavering portrait of his sorrow. The pouring rain plasters his spiky, dark-black hair to his forehead and temples, while glistening water flows down his pale skin, dripping from his chin onto his soaked, black collar. His expression is profound, still and melancholic; his dark, melancholic eyes hold a distant, unfocused gaze, and his lips are set in a soft, neutral line, betraying the deep sorrow within. He remains motionless, like a statue of quiet despair, his only movement the steady stream of rain running down his cheeks and the occasional slow, heavy-lidded blink of his eyes. The scene, bathed in the deep, relentless gloom of the storm but intermittently shattered by bright, cool flashes of lightning, casts his features into stark, high-contrast relief, sculpting the contours of his face in an instant before returning him to the saturated blues and grays of the pouring rain, amplifying the overwhelming atmosphere of quiet despair and isolation.", + "seed": 3890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_4205_seed_3890_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": -0.22070824513211845, + "vlm_logprob": -0.22070824513211845, + "vlm_yes_no_margin": 2.8500000119209288, + "vlm_token_credit": -0.22070824513211845, + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015208004973828793 + ], + "opposite_logprob_tokens": [ + -4.515207767486572 + ], + "logprob_reward": -0.015208004973828793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015208004973828793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47516724467277527, + 0.14394013583660126, + 0.04738779738545418, + 0.06128699332475662, + 0.11842632293701172, + 0.03412254527211189, + 0.04383097216486931, + 0.01966138370335102, + 0.03225509449839592, + 0.023921525105834007 + ], + "credit_map": [ + 0.03353795036673546, + 0.027245037257671356, + 0.13339699804782867, + 0.039929766207933426, + 0.03893328830599785, + 0.12828123569488525, + 0.049901556223630905, + 0.023941408842802048, + 0.010639732703566551, + 0.010181196965277195, + 0.028450604528188705, + 0.010686772875487804, + 0.029575418680906296, + 0.026524944230914116, + 0.017065303400158882, + 0.010816154070198536, + 0.00545879453420639, + 0.005874637980014086, + 0.005353374406695366, + 0.00657747033983469, + 0.0077489325776696205, + 0.005636795423924923, + 0.00657660374417901, + 0.004161188844591379, + 0.005558044649660587, + 0.0077946726232767105, + 0.005966925993561745, + 0.006343873683363199, + 0.013348324224352837, + 0.004901900887489319, + 0.009107248857617378, + 0.008265999145805836, + 0.01388396043330431, + 0.009885547682642937, + 0.014480451121926308, + 0.010681267827749252, + 0.02518349140882492, + 0.009295010007917881, + 0.020287370309233665, + 0.014729222282767296, + 0.004865842871367931, + 0.0035758069716393948, + 0.004368878435343504, + 0.00404237350448966, + 0.004180870950222015, + 0.003565112128853798, + 0.004049093462526798, + 0.005474565550684929, + 0.00796909723430872, + 0.003983251750469208, + 0.006102725863456726, + 0.004780779127031565, + 0.008373432792723179, + 0.004162874072790146, + 0.004462298005819321, + 0.003996510989964008, + 0.0021904774475842714, + 0.0024336096830666065, + 0.0021915773395448923, + 0.002391545567661524, + 0.0032400989439338446, + 0.002059618476778269, + 0.0031481049954891205, + 0.002006351249292493, + 0.0039101215079426765, + 0.003426665673032403, + 0.004130457993596792, + 0.003476733108982444, + 0.006451427936553955, + 0.0031727543100714684, + 0.004409127403050661, + 0.003277806332334876, + 0.0035530910827219486, + 0.002007481874898076, + 0.0026074640918523073, + 0.003020963165909052, + 0.004830485209822655, + 0.0022436438594013453, + 0.0033443965949118137, + 0.002314000390470028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.015208004973828793 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328801512718201 + ], + "opposite_logprob_tokens": [ + -0.7578801512718201 + ], + "logprob_reward": -0.6328801512718201, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328801512718201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29680418968200684, + 0.14270663261413574, + 0.05280536785721779, + 0.08852672576904297, + 0.10438541322946548, + 0.10713537037372589, + 0.05273178964853287, + 0.039102908223867416, + 0.043647754937410355, + 0.07215385138988495 + ], + "credit_map": [ + 0.04390902444720268, + 0.02338724583387375, + 0.07159430533647537, + 0.02062894217669964, + 0.026851516216993332, + 0.055813346058130264, + 0.02697843499481678, + 0.027641374617815018, + 0.019039073958992958, + 0.009052932262420654, + 0.036578256636857986, + 0.007317600306123495, + 0.015137627720832825, + 0.03492189571261406, + 0.009913824498653412, + 0.010745414532721043, + 0.008785133250057697, + 0.004436711315065622, + 0.006964256055653095, + 0.007345679681748152, + 0.007027316372841597, + 0.006656242534518242, + 0.007672042585909367, + 0.003917986527085304, + 0.008667190559208393, + 0.011335095390677452, + 0.009910655207931995, + 0.009144232608377934, + 0.017986956983804703, + 0.008329777047038078, + 0.014142266474664211, + 0.009010547772049904, + 0.008797433227300644, + 0.00910811498761177, + 0.011590990237891674, + 0.013054405339062214, + 0.020528331398963928, + 0.007945484481751919, + 0.021982071921229362, + 0.01137857511639595, + 0.011403389275074005, + 0.014274592511355877, + 0.013055532239377499, + 0.011605730280280113, + 0.016431160271167755, + 0.010394503362476826, + 0.016408279538154602, + 0.013562186621129513, + 0.00866218376904726, + 0.005482396110892296, + 0.007262301631271839, + 0.005472393706440926, + 0.007408943027257919, + 0.006811716128140688, + 0.006415870971977711, + 0.00521598057821393, + 0.006220308598130941, + 0.0037403020542114973, + 0.0044412761926651, + 0.004409413319081068, + 0.005186109337955713, + 0.0039343032985925674, + 0.007210559211671352, + 0.003960636910051107, + 0.006181055214256048, + 0.004629999399185181, + 0.005974967032670975, + 0.004617813974618912, + 0.006071992218494415, + 0.00520342867821455, + 0.006351923104375601, + 0.0046165757812559605, + 0.010289625264704227, + 0.009933617897331715, + 0.007969515398144722, + 0.005894613917917013, + 0.014409986324608326, + 0.007349323946982622, + 0.009442427195608616, + 0.006864737253636122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.6328801512718201 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955924745649099 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.005955924745649099, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955924745649099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42942532896995544, + 0.08936847001314163, + 0.054410215467214584, + 0.04216485098004341, + 0.14657039940357208, + 0.11008859425783157, + 0.04099399223923683, + 0.025108415633440018, + 0.03395811468362808, + 0.027911631390452385 + ], + "credit_map": [ + 0.035963788628578186, + 0.05460822582244873, + 0.06187829002737999, + 0.0495632030069828, + 0.060846298933029175, + 0.0686904713511467, + 0.055811405181884766, + 0.04206365719437599, + 0.0094120679423213, + 0.01037509273737669, + 0.016868064180016518, + 0.00645805848762393, + 0.013950773514807224, + 0.014268342405557632, + 0.010818212293088436, + 0.007217858452349901, + 0.00794135220348835, + 0.00620350381359458, + 0.006204535719007254, + 0.008077345788478851, + 0.009363292716443539, + 0.005639231763780117, + 0.006692300084978342, + 0.004288653843104839, + 0.005103518255054951, + 0.005187679082155228, + 0.005001367069780827, + 0.006464974023401737, + 0.006753704510629177, + 0.0037436755374073982, + 0.006248967722058296, + 0.003660965710878372, + 0.01515747606754303, + 0.01632663793861866, + 0.014311100356280804, + 0.0114351911470294, + 0.034230977296829224, + 0.014343180693686008, + 0.02374415285885334, + 0.0170216653496027, + 0.011877700686454773, + 0.012900556437671185, + 0.013424753211438656, + 0.015208913013339043, + 0.013160762377083302, + 0.01094552781432867, + 0.016138315200805664, + 0.016432063654065132, + 0.005381068214774132, + 0.003474348457530141, + 0.004878922365605831, + 0.004600645042955875, + 0.009244244545698166, + 0.003996619954705238, + 0.005251497961580753, + 0.004166643135249615, + 0.004210236016660929, + 0.0030985872726887465, + 0.0029789283871650696, + 0.002965491497889161, + 0.003483507549390197, + 0.002448918530717492, + 0.0031366716139018536, + 0.0027860754635185003, + 0.004692322574555874, + 0.0028020506724715233, + 0.0036930746864527464, + 0.004723323509097099, + 0.004292459227144718, + 0.0028867009095847607, + 0.006926202680915594, + 0.003941980190575123, + 0.003898901166394353, + 0.003538060700520873, + 0.003404233604669571, + 0.0024360090028494596, + 0.0049841878935694695, + 0.0031472838018089533, + 0.0035054106265306473, + 0.002997542964294553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.005955924745649099 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205826573073864 + ], + "opposite_logprob_tokens": [ + -3.895582675933838 + ], + "logprob_reward": -0.0205826573073864, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205826573073864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3912120461463928, + 0.09654111415147781, + 0.06720887869596481, + 0.06722409278154373, + 0.07481130212545395, + 0.14413198828697205, + 0.03318815305829048, + 0.03675367683172226, + 0.038224510848522186, + 0.05070424824953079 + ], + "credit_map": [ + 0.053338561207056046, + 0.03213394433259964, + 0.07544386386871338, + 0.032840363681316376, + 0.038993898779153824, + 0.06816662102937698, + 0.049632225185632706, + 0.04066257178783417, + 0.012945554219186306, + 0.010733752511441708, + 0.010702623054385185, + 0.007637845817953348, + 0.018134957179427147, + 0.012191010639071465, + 0.012962713837623596, + 0.011232652701437473, + 0.007927309721708298, + 0.007301243022084236, + 0.00865409430116415, + 0.008225282654166222, + 0.011960571631789207, + 0.00803759042173624, + 0.009054528549313545, + 0.00604825746268034, + 0.005701427347958088, + 0.006812623236328363, + 0.007254330441355705, + 0.011792391538619995, + 0.012144203297793865, + 0.0067544858902692795, + 0.009994289837777615, + 0.0067703393287956715, + 0.00819448009133339, + 0.009275177493691444, + 0.008623604662716389, + 0.007049434818327427, + 0.014754509553313255, + 0.006746708881109953, + 0.011040273122489452, + 0.0091271186247468, + 0.013669639825820923, + 0.018100837245583534, + 0.015103799290955067, + 0.020793491974473, + 0.020344994962215424, + 0.014354481361806393, + 0.023637661710381508, + 0.018127083778381348, + 0.004378208424896002, + 0.002916924422606826, + 0.004236655775457621, + 0.0027319607324898243, + 0.005614783614873886, + 0.0037230346351861954, + 0.005870338995009661, + 0.003716249717399478, + 0.0058081564493477345, + 0.003588560037314892, + 0.004641225561499596, + 0.00463902810588479, + 0.004023944493383169, + 0.004046971909701824, + 0.005473629105836153, + 0.004532159771770239, + 0.005902951583266258, + 0.0033587703946977854, + 0.004379092715680599, + 0.005065866746008396, + 0.00452167633920908, + 0.0036968400236219168, + 0.007339342497289181, + 0.00395997054874897, + 0.008112606592476368, + 0.005337425973266363, + 0.006581121124327183, + 0.005044954363256693, + 0.008139998652040958, + 0.0050438642501831055, + 0.006512997671961784, + 0.005931276362389326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0205826573073864 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42891448736190796 + ], + "opposite_logprob_tokens": [ + -1.0539145469665527 + ], + "logprob_reward": -0.42891448736190796, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42891448736190796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5867716073989868, + 0.07062072306871414, + 0.06587052345275879, + 0.04522999748587608, + 0.03493533283472061, + 0.09659609943628311, + 0.02562212571501732, + 0.027663003653287888, + 0.028395215049386024, + 0.018295446410775185 + ], + "credit_map": [ + 0.08448825776576996, + 0.0253257155418396, + 0.1922452747821808, + 0.02656344696879387, + 0.029228776693344116, + 0.16270644962787628, + 0.0319322906434536, + 0.034281350672245026, + 0.010919809341430664, + 0.00645079230889678, + 0.013731173239648342, + 0.00582869490608573, + 0.008183768019080162, + 0.01126034650951624, + 0.00860088411718607, + 0.005645254161208868, + 0.009337013587355614, + 0.006765742786228657, + 0.008862338960170746, + 0.010079311206936836, + 0.00958873052150011, + 0.008931377902626991, + 0.007328614126890898, + 0.004977394826710224, + 0.010382726788520813, + 0.004797097761183977, + 0.0057926722802221775, + 0.005821321625262499, + 0.005752823315560818, + 0.003889846382662654, + 0.004930637311190367, + 0.0038628713227808475, + 0.005931384861469269, + 0.004021477419883013, + 0.004683649633079767, + 0.004088189918547869, + 0.004096198361366987, + 0.0032643661834299564, + 0.004643827676773071, + 0.004206238314509392, + 0.01784820854663849, + 0.00972550269216299, + 0.013590280897915363, + 0.01103546004742384, + 0.011401764117181301, + 0.00966119859367609, + 0.011445702984929085, + 0.011887984350323677, + 0.004478144459426403, + 0.0027305311523377895, + 0.003390026744455099, + 0.002818600507453084, + 0.004086983390152454, + 0.0026829263661056757, + 0.002715121256187558, + 0.0027197925373911858, + 0.0062095168977975845, + 0.002854475751519203, + 0.003818693570792675, + 0.0036830047611147165, + 0.0024486409965902567, + 0.002370959147810936, + 0.003247635904699564, + 0.0030300745274871588, + 0.003181633772328496, + 0.0030810378957539797, + 0.003192546544596553, + 0.003316394053399563, + 0.00502945389598608, + 0.0028354967944324017, + 0.004893443547189236, + 0.002865207614377141, + 0.003982797730714083, + 0.0021248094271868467, + 0.00221540336497128, + 0.0018590862164273858, + 0.0027330864686518908, + 0.0015874276868999004, + 0.0017616802360862494, + 0.0020311553962528706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.42891448736190796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames the face of a young man, Moriya Ishida, caught in the center of a raging thunderstorm. The camera remains completely still, creating an intimate and unwavering portrait of his sorrow. The pouring rain plasters his spiky, dark-black hair to his forehead and temples, while glistening water flows down his pale skin, dripping from his chin onto his soaked, black collar. His expression is profound, still and melancholic; his dark, melancholic eyes hold a distant, unfocused gaze, and his lips are set in a soft, neutral line, betraying the deep sorrow within. He remains motionless, like a statue of quiet despair, his only movement the steady stream of rain running down his cheeks and the occasional slow, heavy-lidded blink of his eyes. The scene, bathed in the deep, relentless gloom of the storm but intermittently shattered by bright, cool flashes of lightning, casts his features into stark, high-contrast relief, sculpting the contours of his face in an instant before returning him to the saturated blues and grays of the pouring rain, amplifying the overwhelming atmosphere of quiet despair and isolation.", + "seed": 3891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_4205_seed_3891_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": -0.22728670975193382, + "vlm_logprob": -0.22728670975193382, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.22728670975193382, + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006835174281150103 + ], + "opposite_logprob_tokens": [ + -5.631834983825684 + ], + "logprob_reward": -0.006835174281150103, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006835174281150103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21680447459220886, + 0.09484568238258362, + 0.13723354041576385, + 0.12701131403446198, + 0.04538338631391525, + 0.08110392093658447, + 0.04843926429748535, + 0.06952761113643646, + 0.14214453101158142, + 0.037506211549043655 + ], + "credit_map": [ + 0.016545932739973068, + 0.02466016262769699, + 0.01545918919146061, + 0.025562606751918793, + 0.027722615748643875, + 0.036991018801927567, + 0.032999277114868164, + 0.036863673478364944, + 0.006851259618997574, + 0.012289274483919144, + 0.01134237740188837, + 0.015524488873779774, + 0.010101823136210442, + 0.013077483512461185, + 0.01564638875424862, + 0.010012587532401085, + 0.013325119391083717, + 0.020571839064359665, + 0.010229158215224743, + 0.019066397100687027, + 0.02386349067091942, + 0.013645906932651997, + 0.02115437388420105, + 0.015377254225313663, + 0.018597962334752083, + 0.01568813994526863, + 0.011341432109475136, + 0.016338909044861794, + 0.012599967420101166, + 0.013659078627824783, + 0.021435430273413658, + 0.017350396141409874, + 0.0034120965283364058, + 0.0059559824876487255, + 0.006947755813598633, + 0.005943194031715393, + 0.004534929059445858, + 0.005991534795612097, + 0.0053322226740419865, + 0.007265673950314522, + 0.004677640274167061, + 0.011138916015625, + 0.008856476284563541, + 0.012665869668126106, + 0.01039428636431694, + 0.008566665463149548, + 0.014112085103988647, + 0.010691980831325054, + 0.0054628364741802216, + 0.008243298158049583, + 0.006052814424037933, + 0.006475445814430714, + 0.0060334536246955395, + 0.0047942595556378365, + 0.005290559027343988, + 0.006086596753448248, + 0.00653128232806921, + 0.008358053863048553, + 0.005579209420830011, + 0.009191647171974182, + 0.00809559691697359, + 0.009942112490534782, + 0.009121449664235115, + 0.012708261609077454, + 0.016715295612812042, + 0.019847029820084572, + 0.011896037496626377, + 0.015284843742847443, + 0.019486377015709877, + 0.02076728455722332, + 0.015424942597746849, + 0.02272270992398262, + 0.0030060771387070417, + 0.005037236027419567, + 0.004655898083001375, + 0.005276210606098175, + 0.00418980373069644, + 0.0037259566597640514, + 0.007136575877666473, + 0.004478453192859888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.006835174281150103 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762175917625427 + ], + "opposite_logprob_tokens": [ + -0.8262175917625427 + ], + "logprob_reward": -0.5762175917625427, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762175917625427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11189871281385422, + 0.08225203305482864, + 0.04310847446322441, + 0.06437817960977554, + 0.04753269627690315, + 0.03432825207710266, + 0.1151159331202507, + 0.17370964586734772, + 0.28925928473472595, + 0.03841671347618103 + ], + "credit_map": [ + 0.015578695572912693, + 0.012509186752140522, + 0.016485201194882393, + 0.013734806329011917, + 0.012948085553944111, + 0.013165390118956566, + 0.014483602717518806, + 0.012993765994906425, + 0.0071702380664646626, + 0.010088255628943443, + 0.008627450093626976, + 0.009815400466322899, + 0.012502577155828476, + 0.010973026044666767, + 0.010349852964282036, + 0.01272523868829012, + 0.0039188070222735405, + 0.0050415946170687675, + 0.006163343787193298, + 0.006196355447173119, + 0.006449251435697079, + 0.0055539775639772415, + 0.004541323985904455, + 0.005243827123194933, + 0.006274824030697346, + 0.008613206446170807, + 0.00579058937728405, + 0.008009793236851692, + 0.007574348244816065, + 0.008284701965749264, + 0.009510085918009281, + 0.010320637375116348, + 0.003564407816156745, + 0.007024696562439203, + 0.005012371111661196, + 0.007778326515108347, + 0.005974915344268084, + 0.00473143020644784, + 0.008275235071778297, + 0.005171316210180521, + 0.0032122284173965454, + 0.004630637355148792, + 0.00418937997892499, + 0.004244242329150438, + 0.005364021752029657, + 0.0044460175558924675, + 0.004814737476408482, + 0.003426993265748024, + 0.019382603466510773, + 0.017319537699222565, + 0.01373476441949606, + 0.012525876984000206, + 0.01574595831334591, + 0.011740587651729584, + 0.01046113669872284, + 0.01420548278838396, + 0.013579154387116432, + 0.020375465974211693, + 0.03032618947327137, + 0.024311378598213196, + 0.02208603173494339, + 0.019664237275719643, + 0.020040765404701233, + 0.023326441645622253, + 0.04109251871705055, + 0.03245457634329796, + 0.041252605617046356, + 0.03168165683746338, + 0.04272128641605377, + 0.0288857389241457, + 0.03167759254574776, + 0.03949332237243652, + 0.0029342486523091793, + 0.005423687864094973, + 0.0037404722534120083, + 0.005674857180565596, + 0.005323686171323061, + 0.004082297906279564, + 0.00682215066626668, + 0.004415315575897694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.5762175917625427 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748740561306477 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748740561306477, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748740561306477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21506400406360626, + 0.12376295775175095, + 0.17156106233596802, + 0.031321752816438675, + 0.07174773514270782, + 0.06465667486190796, + 0.14184020459651947, + 0.0617237314581871, + 0.0974382758140564, + 0.020883606746792793 + ], + "credit_map": [ + 0.02710605226457119, + 0.023055972531437874, + 0.030029065907001495, + 0.023035068064928055, + 0.02771219052374363, + 0.027241401374340057, + 0.027328936383128166, + 0.029555317014455795, + 0.0073236930184066296, + 0.013005136512219906, + 0.01199345476925373, + 0.012316683307290077, + 0.013691216707229614, + 0.027615271508693695, + 0.011246291920542717, + 0.026571208611130714, + 0.013105868361890316, + 0.023544739931821823, + 0.02726871147751808, + 0.027799531817436218, + 0.01990031823515892, + 0.014824438840150833, + 0.026549452915787697, + 0.018567992374300957, + 0.0028750156052410603, + 0.004000145010650158, + 0.003391650505363941, + 0.004700710531324148, + 0.004374877084046602, + 0.0033643185161054134, + 0.00508678425103426, + 0.0035282489843666553, + 0.005455866921693087, + 0.00951921846717596, + 0.01094035618007183, + 0.009678619913756847, + 0.011917266063392162, + 0.005720456596463919, + 0.009430434554815292, + 0.009085522964596748, + 0.003724365495145321, + 0.00831972062587738, + 0.008633499033749104, + 0.010865691117942333, + 0.0063918703235685825, + 0.007125543896108866, + 0.011744762770831585, + 0.007851222530007362, + 0.02259286865592003, + 0.024879375472664833, + 0.011773010715842247, + 0.01737140119075775, + 0.01759117841720581, + 0.015500308945775032, + 0.015250238589942455, + 0.01688181608915329, + 0.0044702705927193165, + 0.008473371155560017, + 0.009552842006087303, + 0.009759078733623028, + 0.0064580924808979034, + 0.007024469319730997, + 0.00817110575735569, + 0.00781449768692255, + 0.016073917970061302, + 0.01257234439253807, + 0.013634903356432915, + 0.010555471293628216, + 0.012091079726815224, + 0.009993675164878368, + 0.00984201580286026, + 0.012674866244196892, + 0.0019048821413889527, + 0.0026678633876144886, + 0.0018303325632587075, + 0.002323415130376816, + 0.0034490854013711214, + 0.002932510105893016, + 0.002705718157812953, + 0.0030697982292622328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.006748740561306477 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291101679205894 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291101679205894, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291101679205894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.162151500582695, + 0.13693809509277344, + 0.1952662169933319, + 0.04035234823822975, + 0.054108668118715286, + 0.051707424223423004, + 0.11859221011400223, + 0.05945459380745888, + 0.14612774550914764, + 0.035301242023706436 + ], + "credit_map": [ + 0.013769598677754402, + 0.015505951829254627, + 0.0172139722853899, + 0.016005627810955048, + 0.01771201193332672, + 0.02573123760521412, + 0.027439739555120468, + 0.028773345053195953, + 0.008247928693890572, + 0.015509279444813728, + 0.014389196410775185, + 0.019187312573194504, + 0.01962687447667122, + 0.017397379502654076, + 0.022008508443832397, + 0.020571598783135414, + 0.011902045458555222, + 0.02374587394297123, + 0.019702009856700897, + 0.024725737050175667, + 0.021713130176067352, + 0.036578789353370667, + 0.018459517508745193, + 0.03843909129500389, + 0.003177829785272479, + 0.004127762280404568, + 0.003027998609468341, + 0.005095724482089281, + 0.005435874219983816, + 0.006481776013970375, + 0.005904738791286945, + 0.007100643590092659, + 0.005205971654504538, + 0.007030994165688753, + 0.0060997409746050835, + 0.00802438985556364, + 0.005240860395133495, + 0.005023466423153877, + 0.011043071746826172, + 0.006440165918320417, + 0.0031533942092210054, + 0.007068385370075703, + 0.005015459377318621, + 0.008914933539927006, + 0.005887897219508886, + 0.005459980573505163, + 0.009896495379507542, + 0.006310874130576849, + 0.018717685714364052, + 0.021900642663240433, + 0.009766904637217522, + 0.014715466648340225, + 0.01650148816406727, + 0.01061090175062418, + 0.012932081706821918, + 0.013447027653455734, + 0.008152333088219166, + 0.008486112579703331, + 0.010332293808460236, + 0.007814189419150352, + 0.005231366492807865, + 0.00640217587351799, + 0.005584150552749634, + 0.007451971527189016, + 0.024253346025943756, + 0.019626351073384285, + 0.01751677878201008, + 0.014893995597958565, + 0.018339281901717186, + 0.016903448849916458, + 0.01429514680057764, + 0.020299382507801056, + 0.0028094928711652756, + 0.005115299019962549, + 0.003557046176865697, + 0.005379280541092157, + 0.003916740883141756, + 0.004086153581738472, + 0.006407693028450012, + 0.004029533825814724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.023291101679205894 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233409404754639 + ], + "opposite_logprob_tokens": [ + -0.8983409404754639 + ], + "logprob_reward": -0.5233409404754639, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233409404754639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25222134590148926, + 0.11536283046007156, + 0.11168326437473297, + 0.0677281990647316, + 0.043404895812273026, + 0.037256523966789246, + 0.06927108019590378, + 0.11045811325311661, + 0.17073245346546173, + 0.02188130095601082 + ], + "credit_map": [ + 0.04093461483716965, + 0.02156912162899971, + 0.037709835916757584, + 0.01939604990184307, + 0.017009560018777847, + 0.04220997169613838, + 0.01902388036251068, + 0.054368309676647186, + 0.008326209150254726, + 0.011918178759515285, + 0.014153126627206802, + 0.013496504165232182, + 0.014834373258054256, + 0.020450206473469734, + 0.014969224110245705, + 0.01721501164138317, + 0.008830295875668526, + 0.011620252393186092, + 0.01595638506114483, + 0.0145396264269948, + 0.011843491345643997, + 0.02080829255282879, + 0.009354361332952976, + 0.018730558454990387, + 0.006475606467574835, + 0.007931403815746307, + 0.007084554526954889, + 0.008722516708076, + 0.007636582013219595, + 0.010216894559562206, + 0.007367089856415987, + 0.012293552048504353, + 0.0038462833035737276, + 0.005263009574264288, + 0.008176601491868496, + 0.005417187698185444, + 0.005507131107151508, + 0.004378547891974449, + 0.005889658350497484, + 0.004926476161926985, + 0.003434377023950219, + 0.004904265981167555, + 0.0048829298466444016, + 0.004733765963464975, + 0.005408089142292738, + 0.004324823152273893, + 0.005421384237706661, + 0.004146886523813009, + 0.00551149807870388, + 0.009690404869616032, + 0.010525933466851711, + 0.008532905019819736, + 0.0072703855112195015, + 0.011827715672552586, + 0.006565290968865156, + 0.009346945211291313, + 0.01615927927196026, + 0.012838237918913364, + 0.014626692980527878, + 0.011709915474057198, + 0.013785123825073242, + 0.014834594912827015, + 0.010685386136174202, + 0.015818893909454346, + 0.02592368982732296, + 0.019127855077385902, + 0.027444681152701378, + 0.019619641825556755, + 0.02239461988210678, + 0.015405796468257904, + 0.018912548199295998, + 0.021903621032834053, + 0.0015336541691794991, + 0.002996983239427209, + 0.0017470549792051315, + 0.003161344211548567, + 0.003165645059198141, + 0.0025653825141489506, + 0.004206957761198282, + 0.002504279837012291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.5233409404754639 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames the face of a young man, Moriya Ishida, caught in the center of a raging thunderstorm. The camera remains completely still, creating an intimate and unwavering portrait of his sorrow. The pouring rain plasters his spiky, dark-black hair to his forehead and temples, while glistening water flows down his pale skin, dripping from his chin onto his soaked, black collar. His expression is profound, still and melancholic; his dark, melancholic eyes hold a distant, unfocused gaze, and his lips are set in a soft, neutral line, betraying the deep sorrow within. He remains motionless, like a statue of quiet despair, his only movement the steady stream of rain running down his cheeks and the occasional slow, heavy-lidded blink of his eyes. The scene, bathed in the deep, relentless gloom of the storm but intermittently shattered by bright, cool flashes of lightning, casts his features into stark, high-contrast relief, sculpting the contours of his face in an instant before returning him to the saturated blues and grays of the pouring rain, amplifying the overwhelming atmosphere of quiet despair and isolation.", + "seed": 3892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_4205_seed_3892_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": -0.22144414968788623, + "vlm_logprob": -0.22144414968788623, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.22144414968788623, + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007906687445938587 + ], + "opposite_logprob_tokens": [ + -5.132906913757324 + ], + "logprob_reward": -0.007906687445938587, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007906687445938587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4095172584056854, + 0.08247508108615875, + 0.09793844819068909, + 0.08717913180589676, + 0.04803720861673355, + 0.0519617460668087, + 0.05289305001497269, + 0.024225492030382156, + 0.12033940106630325, + 0.025433141738176346 + ], + "credit_map": [ + 0.04487486556172371, + 0.04305219277739525, + 0.045541223138570786, + 0.04205784946680069, + 0.04600030183792114, + 0.06220103055238724, + 0.03680510073900223, + 0.08898472785949707, + 0.008081612177193165, + 0.01016986183822155, + 0.009529762901365757, + 0.010538977570831776, + 0.011523048393428326, + 0.009737895801663399, + 0.01145986095070839, + 0.011434074491262436, + 0.009974268265068531, + 0.011171159334480762, + 0.010791696608066559, + 0.011917333118617535, + 0.014306009747087955, + 0.01375066302716732, + 0.011793341487646103, + 0.014233988709747791, + 0.008447136729955673, + 0.013222131878137589, + 0.009056981652975082, + 0.012027120217680931, + 0.0114750312641263, + 0.01254905667155981, + 0.009747058153152466, + 0.01065461803227663, + 0.0062659727409482, + 0.0045648911036551, + 0.00612318841740489, + 0.0065909819677472115, + 0.006713642738759518, + 0.006455826107412577, + 0.005234990734606981, + 0.0060877189971506596, + 0.008852491155266762, + 0.007405367214232683, + 0.006570110563188791, + 0.0065269856713712215, + 0.005407103803008795, + 0.006841999012976885, + 0.004157265182584524, + 0.006200424395501614, + 0.005588840693235397, + 0.006716292817145586, + 0.006393482442945242, + 0.006009944248944521, + 0.008077791891992092, + 0.0064163412898778915, + 0.007410881109535694, + 0.006279478315263987, + 0.0029142301063984632, + 0.0027091726660728455, + 0.002793628955259919, + 0.002960973186418414, + 0.003935419023036957, + 0.0032948660664260387, + 0.0022968901321291924, + 0.003320312825962901, + 0.018274201080203056, + 0.012212553061544895, + 0.013121390715241432, + 0.013459103181958199, + 0.017396707087755203, + 0.020885640755295753, + 0.010723886080086231, + 0.014265920966863632, + 0.0024329060688614845, + 0.0036806983407586813, + 0.002918977290391922, + 0.004604304675012827, + 0.0032094132620841265, + 0.0034037032164633274, + 0.0022398869041353464, + 0.002943252446129918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.007906687445938587 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6934165358543396 + ], + "opposite_logprob_tokens": [ + -0.6934165358543396 + ], + "logprob_reward": -0.6934165358543396, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934165358543396, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4396359920501709, + 0.10606973618268967, + 0.07747935503721237, + 0.06488773971796036, + 0.09909641742706299, + 0.08934763818979263, + 0.05154994875192642, + 0.02544669434428215, + 0.02677925117313862, + 0.019707171246409416 + ], + "credit_map": [ + 0.05530836060643196, + 0.040013864636421204, + 0.05593056231737137, + 0.035118792206048965, + 0.037371400743722916, + 0.06929656863212585, + 0.03856780007481575, + 0.10802866518497467, + 0.011826583184301853, + 0.009340532124042511, + 0.012762478552758694, + 0.009546537883579731, + 0.009451355785131454, + 0.017042750492691994, + 0.008869645185768604, + 0.02722986228764057, + 0.009382779709994793, + 0.0061378744430840015, + 0.011361011303961277, + 0.007994528859853745, + 0.006314501632004976, + 0.013520292937755585, + 0.006148084532469511, + 0.016620293259620667, + 0.008742092177271843, + 0.005888400599360466, + 0.009461085312068462, + 0.006378260441124439, + 0.007795053534209728, + 0.009560303762555122, + 0.005999506916850805, + 0.01106304582208395, + 0.013155569322407246, + 0.010646306909620762, + 0.01310753170400858, + 0.009964254684746265, + 0.015325802378356457, + 0.014721140265464783, + 0.008849944919347763, + 0.013325875625014305, + 0.009938010014593601, + 0.010957730002701283, + 0.011845682747662067, + 0.01486374530941248, + 0.01042371429502964, + 0.009797737933695316, + 0.012096806429326534, + 0.009424219839274883, + 0.007631597574800253, + 0.006268564146012068, + 0.007815795950591564, + 0.007026311941444874, + 0.005739073734730482, + 0.0062180464155972, + 0.005242787767201662, + 0.005607778672128916, + 0.005313157103955746, + 0.002543459879234433, + 0.0037153642624616623, + 0.002603137632831931, + 0.0023398748598992825, + 0.004081521183252335, + 0.0016387266805395484, + 0.00321145448833704, + 0.005734535399824381, + 0.0026576798409223557, + 0.003714501392096281, + 0.0031738150864839554, + 0.0032273579854518175, + 0.0033438922837376595, + 0.002467794343829155, + 0.002459677867591381, + 0.0034263909328728914, + 0.0020150048658251762, + 0.001965839648619294, + 0.0024081701412796974, + 0.0017914031632244587, + 0.0036083562299609184, + 0.001781982951797545, + 0.0027100257575511932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.6934165358543396 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645270295441151 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645270295441151, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645270295441151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38347819447517395, + 0.10016999393701553, + 0.07926003634929657, + 0.09996013343334198, + 0.07352284342050552, + 0.0795748308300972, + 0.06815939396619797, + 0.031827162951231, + 0.020189901813864708, + 0.06385752558708191 + ], + "credit_map": [ + 0.04701124131679535, + 0.035816892981529236, + 0.0446639209985733, + 0.023479171097278595, + 0.02271486073732376, + 0.06422866880893707, + 0.02798621915280819, + 0.11757722496986389, + 0.013269669376313686, + 0.009094111621379852, + 0.012609727680683136, + 0.012975456193089485, + 0.011277570389211178, + 0.015044677071273327, + 0.009497955441474915, + 0.01640082150697708, + 0.0067860521376132965, + 0.008247066289186478, + 0.009733646176755428, + 0.008106281980872154, + 0.00773991271853447, + 0.013458230532705784, + 0.008102357387542725, + 0.017086489126086235, + 0.009967220015823841, + 0.017290154471993446, + 0.007701796479523182, + 0.012232203967869282, + 0.016659006476402283, + 0.0144705381244421, + 0.0112096993252635, + 0.010429520159959793, + 0.012172121554613113, + 0.006706660613417625, + 0.009380348026752472, + 0.00917544960975647, + 0.00799974612891674, + 0.011266650632023811, + 0.006455788388848305, + 0.010366077534854412, + 0.009932705201208591, + 0.01215775404125452, + 0.008732589893043041, + 0.010398388840258121, + 0.012288193218410015, + 0.011037969030439854, + 0.007507507689297199, + 0.007519727107137442, + 0.011582405306398869, + 0.00923861749470234, + 0.007876405492424965, + 0.00802698265761137, + 0.007042688317596912, + 0.009195296093821526, + 0.006149441935122013, + 0.009047554805874825, + 0.004153879825025797, + 0.0040555112063884735, + 0.004374229349195957, + 0.003760674735531211, + 0.004595003090798855, + 0.00416979007422924, + 0.003075497457757592, + 0.003642575116828084, + 0.0036688351538032293, + 0.0022046107333153486, + 0.0025400870945304632, + 0.0021400977857410908, + 0.0020984516013413668, + 0.003152969293296337, + 0.0018342856783419847, + 0.0025505640078336, + 0.006722679827362299, + 0.011606157757341862, + 0.005296808201819658, + 0.009076937101781368, + 0.008749555796384811, + 0.008477605879306793, + 0.005990853998810053, + 0.007936928421258926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.004645270295441151 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30665692687034607, + 0.11392535269260406, + 0.08367525786161423, + 0.10521555691957474, + 0.10991264134645462, + 0.09357578307390213, + 0.05680932477116585, + 0.03896643593907356, + 0.03415029123425484, + 0.05711238831281662 + ], + "credit_map": [ + 0.049265652894973755, + 0.026689158752560616, + 0.05235043540596962, + 0.030701810494065285, + 0.02414843626320362, + 0.04340174049139023, + 0.02294500544667244, + 0.05715468153357506, + 0.016018519178032875, + 0.01094379648566246, + 0.017656361684203148, + 0.009588376618921757, + 0.01287026982754469, + 0.01755361631512642, + 0.011559014208614826, + 0.017735404893755913, + 0.010508320294320583, + 0.007076507434248924, + 0.010343603789806366, + 0.0072431061416864395, + 0.008062594570219517, + 0.01365385577082634, + 0.008117676712572575, + 0.01866959221661091, + 0.012469558976590633, + 0.01305759884417057, + 0.014191684313118458, + 0.009787646122276783, + 0.011429877951741219, + 0.016697250306606293, + 0.010586335323750973, + 0.01699560508131981, + 0.015244402922689915, + 0.009238416329026222, + 0.014669717289507389, + 0.011334788054227829, + 0.014818222261965275, + 0.019986432045698166, + 0.009964123368263245, + 0.014656530693173409, + 0.011117115616798401, + 0.011894234456121922, + 0.013034140691161156, + 0.01414322480559349, + 0.010659183375537395, + 0.012491186149418354, + 0.00934081431478262, + 0.010895886458456516, + 0.012278426438570023, + 0.005491717718541622, + 0.007709895260632038, + 0.007603817153722048, + 0.005028347950428724, + 0.007015761453658342, + 0.005977761000394821, + 0.005703596863895655, + 0.0062631224282085896, + 0.0035476069897413254, + 0.005066100973635912, + 0.004167485982179642, + 0.005232536233961582, + 0.005713704042136669, + 0.0037184988614171743, + 0.005257377866655588, + 0.004303770139813423, + 0.0034299264661967754, + 0.004255984909832478, + 0.00467701954767108, + 0.0037970650009810925, + 0.005345114506781101, + 0.0035372781567275524, + 0.00480413343757391, + 0.00608947454020381, + 0.0062904637306928635, + 0.00780579075217247, + 0.009202348068356514, + 0.006996070966124535, + 0.00769956735894084, + 0.005125532858073711, + 0.007903140969574451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014194712042808533 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870575428009033 + ], + "opposite_logprob_tokens": [ + -1.1370575428009033 + ], + "logprob_reward": -0.3870575428009033, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870575428009033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24500203132629395, + 0.1818670630455017, + 0.08240768313407898, + 0.08722564578056335, + 0.14197133481502533, + 0.08895230293273926, + 0.04034402221441269, + 0.0501721128821373, + 0.025706395506858826, + 0.05635141208767891 + ], + "credit_map": [ + 0.030224090442061424, + 0.016233300790190697, + 0.036154553294181824, + 0.01975996233522892, + 0.017008770257234573, + 0.03933403268456459, + 0.01852458156645298, + 0.06776273250579834, + 0.029473111033439636, + 0.012687724083662033, + 0.02899838797748089, + 0.01666714809834957, + 0.014495205134153366, + 0.02973373606801033, + 0.01786751300096512, + 0.03194423019886017, + 0.009921933524310589, + 0.005764142144471407, + 0.010140057653188705, + 0.00791673269122839, + 0.004812033846974373, + 0.015490582212805748, + 0.008166033774614334, + 0.020196178928017616, + 0.014826484024524689, + 0.008445501327514648, + 0.014708002097904682, + 0.008002348244190216, + 0.01086092833429575, + 0.012054245918989182, + 0.00695746811106801, + 0.011370672844350338, + 0.02427006885409355, + 0.01115870289504528, + 0.020989010110497475, + 0.011929851025342941, + 0.01698223315179348, + 0.026021160185337067, + 0.012024853378534317, + 0.018595458939671516, + 0.014745970256626606, + 0.008351573720574379, + 0.012204318307340145, + 0.010027746669948101, + 0.00955120101571083, + 0.01450742781162262, + 0.009452463127672672, + 0.010111604817211628, + 0.007604374084621668, + 0.004221078474074602, + 0.005350414197891951, + 0.005286551080644131, + 0.00462802080437541, + 0.004801384638994932, + 0.0041491566225886345, + 0.004303042311221361, + 0.008681089617311954, + 0.005019352305680513, + 0.007638633716851473, + 0.005130871199071407, + 0.006595504004508257, + 0.00688862195238471, + 0.0033717912156134844, + 0.0068462505005300045, + 0.0041815717704594135, + 0.0021782887633889914, + 0.0029293482657521963, + 0.002402927028015256, + 0.002529366407543421, + 0.004321316257119179, + 0.003342432901263237, + 0.00382114271633327, + 0.01060029398649931, + 0.004583015106618404, + 0.006950180511921644, + 0.0056055583991110325, + 0.005260510835796595, + 0.01032513752579689, + 0.005337812006473541, + 0.007688901387155056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.3870575428009033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot frames the face of a young man, Moriya Ishida, caught in the center of a raging thunderstorm. The camera remains completely still, creating an intimate and unwavering portrait of his sorrow. The pouring rain plasters his spiky, dark-black hair to his forehead and temples, while glistening water flows down his pale skin, dripping from his chin onto his soaked, black collar. His expression is profound, still and melancholic; his dark, melancholic eyes hold a distant, unfocused gaze, and his lips are set in a soft, neutral line, betraying the deep sorrow within. He remains motionless, like a statue of quiet despair, his only movement the steady stream of rain running down his cheeks and the occasional slow, heavy-lidded blink of his eyes. The scene, bathed in the deep, relentless gloom of the storm but intermittently shattered by bright, cool flashes of lightning, casts his features into stark, high-contrast relief, sculpting the contours of his face in an instant before returning him to the saturated blues and grays of the pouring rain, amplifying the overwhelming atmosphere of quiet despair and isolation.", + "seed": 3893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_11/index_4205_seed_3893_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": -0.25498122815042734, + "vlm_logprob": -0.25498122815042734, + "vlm_yes_no_margin": 2.9499999940395356, + "vlm_token_credit": -0.25498122815042734, + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005321030039340258 + ], + "opposite_logprob_tokens": [ + -6.005321025848389 + ], + "logprob_reward": -0.005321030039340258, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005321030039340258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23368658125400543, + 0.03228068724274635, + 0.022726470604538918, + 0.03154503554105759, + 0.33116886019706726, + 0.047846559435129166, + 0.031597401946783066, + 0.12256673723459244, + 0.11825890839099884, + 0.028322670608758926 + ], + "credit_map": [ + 0.05729322507977486, + 0.012472305446863174, + 0.07857850193977356, + 0.015397739596664906, + 0.01688794232904911, + 0.01980031654238701, + 0.017093760892748833, + 0.016162823885679245, + 0.005842905957251787, + 0.0032996071968227625, + 0.004481946118175983, + 0.004606595728546381, + 0.004540010821074247, + 0.0017206055345013738, + 0.0050561618991196156, + 0.0027328580617904663, + 0.003286460880190134, + 0.002679856726899743, + 0.004127089399844408, + 0.002216041088104248, + 0.002755065681412816, + 0.0018404866568744183, + 0.003928658086806536, + 0.001892813597805798, + 0.0038454991299659014, + 0.003522805869579315, + 0.004436465911567211, + 0.0036725641693919897, + 0.0065680998377501965, + 0.0022793833632022142, + 0.0038863150402903557, + 0.003333904081955552, + 0.08886511623859406, + 0.01745413988828659, + 0.1348329484462738, + 0.016283273696899414, + 0.014685254544019699, + 0.02781822346150875, + 0.01087488979101181, + 0.02035505510866642, + 0.006572302896529436, + 0.0054054115898907185, + 0.005277764052152634, + 0.007047894410789013, + 0.006956400349736214, + 0.004587945062667131, + 0.006929670926183462, + 0.005069175735116005, + 0.0032111944165080786, + 0.003307706443592906, + 0.003558962605893612, + 0.003446789924055338, + 0.004466204904019833, + 0.006252509541809559, + 0.003503478365018964, + 0.0038505576085299253, + 0.016222745180130005, + 0.013964608311653137, + 0.017285101115703583, + 0.015318543650209904, + 0.010792686603963375, + 0.01606326550245285, + 0.01531762070953846, + 0.01760217919945717, + 0.023214159533381462, + 0.00720328139141202, + 0.011991249397397041, + 0.006694181822240353, + 0.006687300279736519, + 0.035890042781829834, + 0.011697495356202126, + 0.014881201088428497, + 0.0030406347941607237, + 0.003009719541296363, + 0.00374137912876904, + 0.0029746543150395155, + 0.0032902471721172333, + 0.0035091883037239313, + 0.0048405262641608715, + 0.003916323184967041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.005321030039340258 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578722834587097 + ], + "opposite_logprob_tokens": [ + -0.6328722834587097 + ], + "logprob_reward": -0.7578722834587097, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578722834587097, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1506885290145874, + 0.030783088877797127, + 0.025238268077373505, + 0.037493593990802765, + 0.2064322680234909, + 0.08409677445888519, + 0.03351736068725586, + 0.20819827914237976, + 0.14942426979541779, + 0.07412751019001007 + ], + "credit_map": [ + 0.034290771931409836, + 0.01066831685602665, + 0.04386188089847565, + 0.01623474806547165, + 0.011936516501009464, + 0.008668479509651661, + 0.012517839670181274, + 0.012509970925748348, + 0.0037514781579375267, + 0.004076388198882341, + 0.00447792699560523, + 0.003564519342035055, + 0.004860422108322382, + 0.002584022469818592, + 0.004549735691398382, + 0.002918593119829893, + 0.003343794494867325, + 0.002859105123206973, + 0.003618886461481452, + 0.0023394639138132334, + 0.0038904377724975348, + 0.002383812330663204, + 0.003922470845282078, + 0.002880294807255268, + 0.007438368629664183, + 0.0034369565546512604, + 0.005376104731112719, + 0.0033830672036856413, + 0.003818973433226347, + 0.004825705196708441, + 0.004873507656157017, + 0.00434090755879879, + 0.049772318452596664, + 0.011657237075269222, + 0.07988432049751282, + 0.012968417257070541, + 0.012266022153198719, + 0.019120167940855026, + 0.009805435314774513, + 0.010958338156342506, + 0.012262202799320221, + 0.007005997002124786, + 0.009689897298812866, + 0.00985085591673851, + 0.010814414359629154, + 0.016120484098792076, + 0.009673788212239742, + 0.008679130114614964, + 0.0039031365886330605, + 0.0037116955500096083, + 0.003942968789488077, + 0.005018271040171385, + 0.004610272124409676, + 0.0032942460384219885, + 0.004988487809896469, + 0.004048280417919159, + 0.014422891661524773, + 0.012154762633144855, + 0.0578186959028244, + 0.013303588144481182, + 0.014235036447644234, + 0.05859529227018356, + 0.019293969497084618, + 0.01837405003607273, + 0.016339203342795372, + 0.013001459650695324, + 0.018852494657039642, + 0.01337624341249466, + 0.02027716115117073, + 0.031925104558467865, + 0.02259133569896221, + 0.013061270117759705, + 0.011983579955995083, + 0.0071997130289673805, + 0.011114932596683502, + 0.0071080997586250305, + 0.010501869022846222, + 0.009040125645697117, + 0.008166921325027943, + 0.009012263268232346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.7578722834587097 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076352194882929325 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.0076352194882929325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076352194882929325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21814557909965515, + 0.08761285990476608, + 0.050720229744911194, + 0.0370841808617115, + 0.20669101178646088, + 0.07127691805362701, + 0.060721371322870255, + 0.10467196255922318, + 0.0543397031724453, + 0.10873628407716751 + ], + "credit_map": [ + 0.030431455001235008, + 0.019768862053751945, + 0.0288846455514431, + 0.02219241112470627, + 0.03550019487738609, + 0.024940073490142822, + 0.03518204018473625, + 0.021245868876576424, + 0.009960678406059742, + 0.010477219708263874, + 0.01102373655885458, + 0.008563737384974957, + 0.01723085716366768, + 0.006191379390656948, + 0.016920490190386772, + 0.007244753185659647, + 0.006670299451798201, + 0.006119290366768837, + 0.00716940313577652, + 0.0044508809223771095, + 0.010077457875013351, + 0.0032207851763814688, + 0.008559870533645153, + 0.004452235996723175, + 0.004771093837916851, + 0.0045281266793608665, + 0.005284722428768873, + 0.00483882799744606, + 0.0051727318204939365, + 0.0026376156602054834, + 0.0062690116465091705, + 0.003582043806090951, + 0.03575468808412552, + 0.018413400277495384, + 0.05097995698451996, + 0.016584062948822975, + 0.031128592789173126, + 0.018472837284207344, + 0.022822987288236618, + 0.012534460052847862, + 0.011067373678088188, + 0.009136478416621685, + 0.007577755022794008, + 0.006547061260789633, + 0.013559686951339245, + 0.00907247606664896, + 0.008375716395676136, + 0.005940360948443413, + 0.0064130923710763454, + 0.007915963418781757, + 0.007929072715342045, + 0.009840940125286579, + 0.00863683596253395, + 0.004800344817340374, + 0.008067712187767029, + 0.007117403205484152, + 0.009531007148325443, + 0.013366972096264362, + 0.017549345269799232, + 0.010174920782446861, + 0.013066915795207024, + 0.016734862700104713, + 0.014803175814449787, + 0.009444755502045155, + 0.005672024097293615, + 0.004541696980595589, + 0.0065790750086307526, + 0.004792580846697092, + 0.009105215780436993, + 0.009734205901622772, + 0.008647569455206394, + 0.005267326254397631, + 0.017704129219055176, + 0.010222646407783031, + 0.016961941495537758, + 0.009496735408902168, + 0.011051657609641552, + 0.019462307915091515, + 0.010104076005518436, + 0.013732774183154106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799247160553932 + ], + "opposite_logprob_tokens": [ + -3.529799222946167 + ], + "logprob_reward": -0.029799247160553932, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799247160553932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20247097313404083, + 0.06982140243053436, + 0.05220578610897064, + 0.051510244607925415, + 0.22686047852039337, + 0.09486079216003418, + 0.06155167147517204, + 0.06026751548051834, + 0.07156356424093246, + 0.10888753086328506 + ], + "credit_map": [ + 0.023069217801094055, + 0.01787358894944191, + 0.03185040503740311, + 0.03190287575125694, + 0.02534695342183113, + 0.016510294750332832, + 0.03094998188316822, + 0.024967677891254425, + 0.008814438246190548, + 0.008351440541446209, + 0.00839220266789198, + 0.008034007623791695, + 0.011196152307093143, + 0.005886887200176716, + 0.01257737074047327, + 0.006568917538970709, + 0.006694361101835966, + 0.006430733948945999, + 0.00701297540217638, + 0.006126605439931154, + 0.0072419364005327225, + 0.00401788717135787, + 0.009863351471722126, + 0.004817938897758722, + 0.006839555688202381, + 0.006425959523767233, + 0.007315274327993393, + 0.005757797975093126, + 0.00885788258165121, + 0.003848064225167036, + 0.008157849311828613, + 0.004307867027819157, + 0.03873186558485031, + 0.020896917209029198, + 0.04526565968990326, + 0.02175086736679077, + 0.03133545070886612, + 0.024776512756943703, + 0.025780778378248215, + 0.018322443589568138, + 0.018178684636950493, + 0.007437991909682751, + 0.0106570515781641, + 0.008038260973989964, + 0.008710134774446487, + 0.024900980293750763, + 0.007136233150959015, + 0.0098014697432518, + 0.0070458767004311085, + 0.00809051189571619, + 0.006495050620287657, + 0.007727550342679024, + 0.010302448645234108, + 0.005110433790832758, + 0.009560604579746723, + 0.0072191995568573475, + 0.004642913118004799, + 0.005837036296725273, + 0.013590916991233826, + 0.005209429655224085, + 0.0062352134846150875, + 0.013805135153234005, + 0.005409024190157652, + 0.0055378517135977745, + 0.01025693491101265, + 0.005830794107168913, + 0.007951145060360432, + 0.005701378453522921, + 0.007757524028420448, + 0.016767309978604317, + 0.00929640606045723, + 0.008002081885933876, + 0.01900985650718212, + 0.008877873420715332, + 0.01830051839351654, + 0.009047724306583405, + 0.010120407678186893, + 0.01868775300681591, + 0.011658099479973316, + 0.013185306452214718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.029799247160553932 + }, + { + "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 framing the face of a young man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's hair spiky, dark-black, and plastered to his forehead and temples by rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character wearing a soaked black collar with water dripping from his chin onto it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature intermittent bright flashes of lightning amidst the gloom?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47427836060523987 + ], + "opposite_logprob_tokens": [ + -0.9742783308029175 + ], + "logprob_reward": -0.47427836060523987, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47427836060523987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1358424425125122, + 0.040387026965618134, + 0.04826618731021881, + 0.05552561581134796, + 0.16066475212574005, + 0.05309053137898445, + 0.05049770325422287, + 0.0434502437710762, + 0.148143008351326, + 0.2641324996948242 + ], + "credit_map": [ + 0.02469143085181713, + 0.008886841125786304, + 0.03495492786169052, + 0.017070163041353226, + 0.010708645917475224, + 0.01382383517920971, + 0.013162513263523579, + 0.012544074095785618, + 0.006167175248265266, + 0.004508752841502428, + 0.006658283527940512, + 0.005753119941800833, + 0.004349065013229847, + 0.002878919942304492, + 0.005345401354134083, + 0.004726310260593891, + 0.009312489069998264, + 0.00596999516710639, + 0.007038740906864405, + 0.0054299188777804375, + 0.005020801909267902, + 0.003907479345798492, + 0.006033545825630426, + 0.0055532148107886314, + 0.01224619708955288, + 0.005008592735975981, + 0.008788824081420898, + 0.005309741944074631, + 0.005831744056195021, + 0.005866944789886475, + 0.006890328135341406, + 0.005583245772868395, + 0.04002954065799713, + 0.009064466692507267, + 0.06383290886878967, + 0.008864629082381725, + 0.008655660785734653, + 0.014293876476585865, + 0.006520611234009266, + 0.00940305832773447, + 0.008459880016744137, + 0.0048127067275345325, + 0.006747005041688681, + 0.005235908553004265, + 0.00510519091039896, + 0.010494325309991837, + 0.006235525012016296, + 0.00599999213591218, + 0.005826535634696484, + 0.00497196801006794, + 0.005957755725830793, + 0.007898491807281971, + 0.006546670105308294, + 0.0057635377161204815, + 0.00569288432598114, + 0.007839858531951904, + 0.006176934111863375, + 0.005498789716511965, + 0.007361501455307007, + 0.005334040615707636, + 0.004728312138468027, + 0.004199514631181955, + 0.005186718888580799, + 0.004964432679116726, + 0.029304493218660355, + 0.007523976266384125, + 0.01669539138674736, + 0.008617863059043884, + 0.009242271073162556, + 0.04984556883573532, + 0.00796671025454998, + 0.018946737051010132, + 0.051948193460702896, + 0.02419508621096611, + 0.03960070013999939, + 0.024884313344955444, + 0.025985099375247955, + 0.04113641753792763, + 0.024432053789496422, + 0.031950630247592926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.47427836060523987 + } + ], + "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/0000060/group_12/index_tensor([6324, 6324])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_tensor([6324, 6324])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6e5cf580c93969d537d328a5fcaa308f6361c32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_tensor([6324, 6324])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera tilts downward at high speed, captured in extreme slow motion, closely following a crystalline, conical icicle detached from a high rooftop. Encased within tiny bubbles and frozen veins, its mirror-like surface reflects the chilly, clear winter morning light. Plummeting under the weight of gravity, it carves a blurry path through the air, careening towards the dark, frost-covered stone floor below. Upon impact, the icicle's core suddenly explodes, shattering in all directions into countless angular, crystalline fragments of varying sizes. Each fragment, like a miniature diamond, is clearly visible in the slow motion, spinning and tumbling wildly, capturing and refracting the cold morning light. Against the dark background, it creates a fleeting, dazzling haze of light that eventually disperses like stardust. The image is imbued with a pure physical beauty and a chilling, fragile tranquility, dominated by icy blue, dark gray, and high-gloss white.", + "seed": 24001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_6324_seed_24001_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": -0.08353651160141454, + "vlm_logprob": -0.08353651160141454, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.08353651160141454, + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026525801047682762 + ], + "opposite_logprob_tokens": [ + -5.151525974273682 + ], + "logprob_reward": -0.026525801047682762, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.026525801047682762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35890352725982666, + 0.12692993879318237, + 0.1700654923915863, + 0.023650716990232468, + 0.04212920367717743, + 0.05478980019688606, + 0.04216032847762108, + 0.11095263063907623, + 0.03379185125231743, + 0.03662647306919098 + ], + "credit_map": [ + 0.06023509427905083, + 0.07050526142120361, + 0.04455406963825226, + 0.01624753698706627, + 0.019765615463256836, + 0.09642725437879562, + 0.019792791455984116, + 0.03137590363621712, + 0.00920376181602478, + 0.018121900036931038, + 0.019527049735188484, + 0.01748783513903618, + 0.01230345293879509, + 0.011714260093867779, + 0.016913993284106255, + 0.021657688543200493, + 0.01403867732733488, + 0.010816938243806362, + 0.042867641896009445, + 0.015527985990047455, + 0.012018582783639431, + 0.018781689926981926, + 0.018016541376709938, + 0.037997446954250336, + 0.0028592885937541723, + 0.002954015741124749, + 0.0018391440389677882, + 0.0024284240789711475, + 0.005466262809932232, + 0.003004322061315179, + 0.002629195572808385, + 0.002470065839588642, + 0.003981088753789663, + 0.004464445635676384, + 0.007905581966042519, + 0.007804587949067354, + 0.004397140350192785, + 0.0034631823655217886, + 0.004426045808941126, + 0.005687137134373188, + 0.005600336007773876, + 0.0072333370335400105, + 0.007672375533729792, + 0.0092035923153162, + 0.004496487323194742, + 0.004688129760324955, + 0.009299700148403645, + 0.0065958439372479916, + 0.003795326454564929, + 0.0042776078917086124, + 0.005322579760104418, + 0.005646900739520788, + 0.0037284826394170523, + 0.006535903085023165, + 0.005435362923890352, + 0.007418164983391762, + 0.005941272713243961, + 0.004381559323519468, + 0.039061322808265686, + 0.007813802920281887, + 0.005155863240361214, + 0.006647375412285328, + 0.011088155210018158, + 0.030863288789987564, + 0.004178316332399845, + 0.0037276407238096, + 0.004841164220124483, + 0.00469091534614563, + 0.004576352424919605, + 0.0035006708931177855, + 0.0038855329621583223, + 0.0043912590481340885, + 0.005227969028055668, + 0.003335100132972002, + 0.004689265042543411, + 0.006321784574538469, + 0.0029630016069859266, + 0.002930163173004985, + 0.004309774376451969, + 0.006849418394267559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026525801047682762 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870505094528198 + ], + "opposite_logprob_tokens": [ + -1.1370505094528198 + ], + "logprob_reward": -0.3870505094528198, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870505094528198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19394522905349731, + 0.1352630853652954, + 0.21450765430927277, + 0.037054114043712616, + 0.07196926325559616, + 0.06287748366594315, + 0.11797305941581726, + 0.05758767947554588, + 0.08697126805782318, + 0.021851178258657455 + ], + "credit_map": [ + 0.03443213552236557, + 0.027436479926109314, + 0.03207006677985191, + 0.014085586182773113, + 0.013997409492731094, + 0.038366422057151794, + 0.012560482136905193, + 0.020996637642383575, + 0.016232993453741074, + 0.020124610513448715, + 0.02428986318409443, + 0.014026453718543053, + 0.013632775284349918, + 0.017085932195186615, + 0.013813420198857784, + 0.016057025641202927, + 0.014159243553876877, + 0.00628798920661211, + 0.08591629564762115, + 0.015199494548141956, + 0.005895846523344517, + 0.008841308765113354, + 0.014227231964468956, + 0.06398025155067444, + 0.004292219411581755, + 0.004166648257523775, + 0.0036643685307353735, + 0.004089150112122297, + 0.005809335503727198, + 0.005362323019653559, + 0.0051086959429085255, + 0.0045613739639520645, + 0.0065037552267313, + 0.005637481343001127, + 0.011273284442722797, + 0.01269337348639965, + 0.013184825889766216, + 0.005892544519156218, + 0.00821700133383274, + 0.008567003533244133, + 0.007129227742552757, + 0.011884305626153946, + 0.013540314510464668, + 0.008568946272134781, + 0.005216311663389206, + 0.003418068401515484, + 0.007429077289998531, + 0.005691234953701496, + 0.01008431427180767, + 0.011785607784986496, + 0.019517328590154648, + 0.024826722219586372, + 0.008621519431471825, + 0.012174318544566631, + 0.015662536025047302, + 0.015300706960260868, + 0.004486445803195238, + 0.005885162390768528, + 0.007908101193606853, + 0.0065165068954229355, + 0.006965331267565489, + 0.007476134691387415, + 0.009445643052458763, + 0.008904354646801949, + 0.012310070917010307, + 0.011978053487837315, + 0.01017268467694521, + 0.00956667773425579, + 0.011345738545060158, + 0.01381910964846611, + 0.010370239615440369, + 0.0074086906388401985, + 0.0027220312040299177, + 0.003394237719476223, + 0.0023317753802984953, + 0.0030298454221338034, + 0.0026660419534891844, + 0.0022572833113372326, + 0.002595151076093316, + 0.002854812890291214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.3870505094528198 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019524098606780171 + ], + "opposite_logprob_tokens": [ + -6.251952648162842 + ], + "logprob_reward": -0.0019524098606780171, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019524098606780171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1343827247619629, + 0.09123779088258743, + 0.3337012827396393, + 0.02967694029211998, + 0.06635542958974838, + 0.11018793284893036, + 0.05353621765971184, + 0.08907976746559143, + 0.04447098448872566, + 0.04737091064453125 + ], + "credit_map": [ + 0.016195083037018776, + 0.023705866187810898, + 0.016056304797530174, + 0.010342925786972046, + 0.011316167190670967, + 0.0302857868373394, + 0.012869262136518955, + 0.013611342757940292, + 0.007744406815618277, + 0.01433099526911974, + 0.011309154331684113, + 0.010937721468508244, + 0.009601199068129063, + 0.010658640414476395, + 0.014305617660284042, + 0.012350063771009445, + 0.02040771208703518, + 0.026752158999443054, + 0.08656736463308334, + 0.029146423563361168, + 0.0240738857537508, + 0.031168580055236816, + 0.032877106219530106, + 0.08270805329084396, + 0.0030771277379244566, + 0.003090368816629052, + 0.002945642452687025, + 0.0033128366339951754, + 0.005380107555538416, + 0.004835081752389669, + 0.004048299975693226, + 0.002987476298585534, + 0.008901906199753284, + 0.00821496918797493, + 0.0061582159250974655, + 0.0050592366605997086, + 0.01711234264075756, + 0.006841311696916819, + 0.007527088280767202, + 0.006540367379784584, + 0.009876139461994171, + 0.012189537286758423, + 0.0180143304169178, + 0.022279931232333183, + 0.008297554217278957, + 0.006539070513099432, + 0.02080521360039711, + 0.01218616683036089, + 0.005131907761096954, + 0.00506934616714716, + 0.007266412489116192, + 0.006446389481425285, + 0.004791323561221361, + 0.009120086207985878, + 0.005753497127443552, + 0.009957258589565754, + 0.003913078457117081, + 0.004594852682203054, + 0.03130936622619629, + 0.0064752972684800625, + 0.005839459132403135, + 0.006000904366374016, + 0.010342678055167198, + 0.02060413360595703, + 0.005489107221364975, + 0.005245202220976353, + 0.006219063885509968, + 0.0054206280037760735, + 0.005472666583955288, + 0.0057104527950286865, + 0.00617759907618165, + 0.004736270289868116, + 0.003960400354117155, + 0.004045031499117613, + 0.005667342338711023, + 0.008208252489566803, + 0.0054856715723872185, + 0.004472293425351381, + 0.007899509742856026, + 0.007632415276020765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0019524098606780171 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2514910399913788, + 0.06980828195810318, + 0.21879175305366516, + 0.028672916814684868, + 0.11089592427015305, + 0.1586543172597885, + 0.04695413261651993, + 0.048620354384183884, + 0.04626576229929924, + 0.019845495000481606 + ], + "credit_map": [ + 0.030056387186050415, + 0.033478427678346634, + 0.030171561986207962, + 0.024341247975826263, + 0.032915838062763214, + 0.04705571010708809, + 0.02902265265583992, + 0.024449223652482033, + 0.005815635900944471, + 0.010213561356067657, + 0.008645737543702126, + 0.008869882673025131, + 0.008072573691606522, + 0.007304835598915815, + 0.009979680180549622, + 0.010906376875936985, + 0.01477748341858387, + 0.012534924782812595, + 0.06544866412878036, + 0.01654842309653759, + 0.013721480965614319, + 0.01788341999053955, + 0.020744530484080315, + 0.057132817804813385, + 0.0036214846186339855, + 0.003471380565315485, + 0.002780080307275057, + 0.0029801432974636555, + 0.005242296028882265, + 0.0035907512065023184, + 0.0037787186447530985, + 0.003208060748875141, + 0.015161557123064995, + 0.017473900690674782, + 0.013977938331663609, + 0.012781471945345402, + 0.011415436863899231, + 0.008669165894389153, + 0.02057816833257675, + 0.010838278569281101, + 0.01249391958117485, + 0.015289241448044777, + 0.035397306084632874, + 0.026256440207362175, + 0.009240261279046535, + 0.007857311517000198, + 0.029487984254956245, + 0.022631842643022537, + 0.005628819111734629, + 0.004737100098282099, + 0.007360219489783049, + 0.006425575353205204, + 0.0050821928307414055, + 0.0038333332631736994, + 0.006360105238854885, + 0.007526783272624016, + 0.005407184362411499, + 0.003270006738603115, + 0.011335075832903385, + 0.005730769131332636, + 0.002814091509208083, + 0.004597882740199566, + 0.009732804261147976, + 0.00573253957554698, + 0.007024996913969517, + 0.006938972044736147, + 0.005454347003251314, + 0.004609756171703339, + 0.005847541615366936, + 0.006830253638327122, + 0.004967218264937401, + 0.004592675715684891, + 0.0032097299117594957, + 0.003265372710302472, + 0.0021445895545184612, + 0.0020167408511042595, + 0.002200251678004861, + 0.00216020829975605, + 0.0024037081748247147, + 0.0024448942858725786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0013371107634156942 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008167268824763596 + ], + "opposite_logprob_tokens": [ + -7.125816822052002 + ], + "logprob_reward": -0.0008167268824763596, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008167268824763596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15281963348388672, + 0.053819701075553894, + 0.3195599615573883, + 0.03228646516799927, + 0.06990084052085876, + 0.12433842569589615, + 0.03937867283821106, + 0.1131363958120346, + 0.052915021777153015, + 0.04184485226869583 + ], + "credit_map": [ + 0.02254347689449787, + 0.02023584395647049, + 0.030560795217752457, + 0.010580379515886307, + 0.010539371520280838, + 0.026628777384757996, + 0.009853082709014416, + 0.021877901628613472, + 0.005067499820142984, + 0.007242256309837103, + 0.007497523911297321, + 0.006847580894827843, + 0.006011724006384611, + 0.004808456636965275, + 0.007059973198920488, + 0.009284686297178268, + 0.018621891736984253, + 0.01747719570994377, + 0.11443299800157547, + 0.01807948760688305, + 0.011586408130824566, + 0.015340124256908894, + 0.019375422969460487, + 0.10464643687009811, + 0.0039321305230259895, + 0.0036072672810405493, + 0.0039013270288705826, + 0.003510948270559311, + 0.005787078756839037, + 0.004485714249312878, + 0.003996275365352631, + 0.0030657239258289337, + 0.008209513500332832, + 0.009266018867492676, + 0.008731840178370476, + 0.007824431173503399, + 0.011681939475238323, + 0.005946113262325525, + 0.01135463360697031, + 0.006886350456625223, + 0.011815468780696392, + 0.013797524385154247, + 0.023746874183416367, + 0.021272702142596245, + 0.00847456231713295, + 0.008809765800833702, + 0.021875515580177307, + 0.014546017162501812, + 0.0036797518841922283, + 0.0034103707876056433, + 0.004602302331477404, + 0.0053770216181874275, + 0.003772060154005885, + 0.005984483286738396, + 0.004694125149399042, + 0.007858557626605034, + 0.0057023693807423115, + 0.0046891337260603905, + 0.04235085844993591, + 0.007290347013622522, + 0.0035586224403232336, + 0.004991146270185709, + 0.010810758918523788, + 0.03374315798282623, + 0.007648273836821318, + 0.007294957526028156, + 0.008384660817682743, + 0.005544634070247412, + 0.005851529538631439, + 0.00677981972694397, + 0.0068685393780469894, + 0.004542605020105839, + 0.004419511649757624, + 0.0055367606692016125, + 0.005167983938008547, + 0.006766641978174448, + 0.004941004794090986, + 0.0037278670351952314, + 0.005248036701232195, + 0.006037044804543257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0008167268824763596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera tilts downward at high speed, captured in extreme slow motion, closely following a crystalline, conical icicle detached from a high rooftop. Encased within tiny bubbles and frozen veins, its mirror-like surface reflects the chilly, clear winter morning light. Plummeting under the weight of gravity, it carves a blurry path through the air, careening towards the dark, frost-covered stone floor below. Upon impact, the icicle's core suddenly explodes, shattering in all directions into countless angular, crystalline fragments of varying sizes. Each fragment, like a miniature diamond, is clearly visible in the slow motion, spinning and tumbling wildly, capturing and refracting the cold morning light. Against the dark background, it creates a fleeting, dazzling haze of light that eventually disperses like stardust. The image is imbued with a pure physical beauty and a chilling, fragile tranquility, dominated by icy blue, dark gray, and high-gloss white.", + "seed": 24002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_6324_seed_24002_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": -0.034634044824633746, + "vlm_logprob": -0.034634044824633746, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.034634044824633746, + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009345239959657192 + ], + "opposite_logprob_tokens": [ + -6.384345054626465 + ], + "logprob_reward": -0.009345239959657192, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.009345239959657192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19964361190795898, + 0.05216583237051964, + 0.18953979015350342, + 0.11005251854658127, + 0.07777152210474014, + 0.04599113017320633, + 0.10396318882703781, + 0.09919340163469315, + 0.05935114622116089, + 0.062327854335308075 + ], + "credit_map": [ + 0.021427566185593605, + 0.010772706009447575, + 0.038920965045690536, + 0.02263261377811432, + 0.02554456703364849, + 0.024881863966584206, + 0.03029663674533367, + 0.025166690349578857, + 0.0048746694810688496, + 0.005536193028092384, + 0.006617124658077955, + 0.007386331912130117, + 0.008213547989726067, + 0.007360442541539669, + 0.008249619044363499, + 0.00392790324985981, + 0.01993592642247677, + 0.009829831309616566, + 0.056906457990407944, + 0.009683169424533844, + 0.009752103127539158, + 0.009901082143187523, + 0.013563134707510471, + 0.05996808409690857, + 0.009948424994945526, + 0.016745682805776596, + 0.01492924802005291, + 0.01635550707578659, + 0.017217937856912613, + 0.011910736560821533, + 0.014253659173846245, + 0.008691315539181232, + 0.005981208756566048, + 0.007382807321846485, + 0.004953472875058651, + 0.0064358399249613285, + 0.011819876730442047, + 0.01852429285645485, + 0.008317696861922741, + 0.01435632910579443, + 0.005638415925204754, + 0.005909254774451256, + 0.0066840010695159435, + 0.006167255342006683, + 0.004661583807319403, + 0.004694397561252117, + 0.007767968345433474, + 0.004468252416700125, + 0.016423087567090988, + 0.007103980518877506, + 0.007729978766292334, + 0.01889113150537014, + 0.006085712928324938, + 0.005463100969791412, + 0.036987245082855225, + 0.0052789486944675446, + 0.006352950818836689, + 0.012971178628504276, + 0.008508188650012016, + 0.018278537318110466, + 0.011471346952021122, + 0.007884319871664047, + 0.023547906428575516, + 0.010178966447710991, + 0.004396728705614805, + 0.006790817715227604, + 0.006614222656935453, + 0.005068721249699593, + 0.021016372367739677, + 0.006322928238660097, + 0.005162654910236597, + 0.003978699445724487, + 0.005178291350603104, + 0.00981661956757307, + 0.006824415642768145, + 0.009087084792554379, + 0.014654907397925854, + 0.003999564331024885, + 0.008752891793847084, + 0.004014084581285715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.009345239959657192 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034261882305145 + ], + "opposite_logprob_tokens": [ + -1.9103425741195679 + ], + "logprob_reward": -0.16034261882305145, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034261882305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20986372232437134, + 0.06222624331712723, + 0.11501119285821915, + 0.12072091549634933, + 0.042272429913282394, + 0.09427914023399353, + 0.17735235393047333, + 0.10905471444129944, + 0.043732330203056335, + 0.025486968457698822 + ], + "credit_map": [ + 0.020138269290328026, + 0.019007645547389984, + 0.03479349613189697, + 0.026110583916306496, + 0.02748441882431507, + 0.025801414623856544, + 0.02574743516743183, + 0.030780453234910965, + 0.00628329161554575, + 0.004964035004377365, + 0.00985727645456791, + 0.011938679032027721, + 0.009384221397340298, + 0.007501959800720215, + 0.006918721832334995, + 0.005378056317567825, + 0.012160750105977058, + 0.0077447653748095036, + 0.028146356344223022, + 0.007069520652294159, + 0.00841837003827095, + 0.010426783002912998, + 0.009830563329160213, + 0.031214095652103424, + 0.00879861693829298, + 0.026475939899683, + 0.014502638019621372, + 0.014050348661839962, + 0.017850268632173538, + 0.010129202157258987, + 0.014309375546872616, + 0.014604524709284306, + 0.006899903994053602, + 0.0044267321936786175, + 0.00434662913903594, + 0.0036734973546117544, + 0.0077750845812261105, + 0.004049442708492279, + 0.006650274153798819, + 0.004450868349522352, + 0.009703936986625195, + 0.010739943943917751, + 0.01015983335673809, + 0.010856082662940025, + 0.021005716174840927, + 0.012903518043458462, + 0.01053893007338047, + 0.008371178060770035, + 0.02014492079615593, + 0.012878813780844212, + 0.006320463493466377, + 0.034867800772190094, + 0.012171903625130653, + 0.009266579523682594, + 0.07166623324155807, + 0.010035652667284012, + 0.007290549110621214, + 0.022505972534418106, + 0.0121541079133749, + 0.01831183210015297, + 0.013024729676544666, + 0.010364403016865253, + 0.014820706099271774, + 0.010582415387034416, + 0.0034310552291572094, + 0.004855704959481955, + 0.006024868693202734, + 0.005239400546997786, + 0.012681661173701286, + 0.003616987494751811, + 0.005193336401134729, + 0.0026893126778304577, + 0.002852880861610174, + 0.003271527122706175, + 0.002721846802160144, + 0.004232601262629032, + 0.004376961383968592, + 0.002026564907282591, + 0.004136290866881609, + 0.00186829490121454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.16034261882305145 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015197168104350567 + ], + "opposite_logprob_tokens": [ + -6.501519680023193 + ], + "logprob_reward": -0.0015197168104350567, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015197168104350567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22200807929039001, + 0.05993358790874481, + 0.05326598510146141, + 0.2569573223590851, + 0.06010008603334427, + 0.10534432530403137, + 0.13646481931209564, + 0.03623579069972038, + 0.039110150188207626, + 0.03057987056672573 + ], + "credit_map": [ + 0.022530464455485344, + 0.020143674686551094, + 0.027131786569952965, + 0.028375081717967987, + 0.03295428305864334, + 0.028012938797473907, + 0.030764427036046982, + 0.03209541738033295, + 0.006459968164563179, + 0.006083985790610313, + 0.005917578935623169, + 0.007272400893270969, + 0.009785276837646961, + 0.007032633293420076, + 0.01029566116631031, + 0.007086085621267557, + 0.005377073306590319, + 0.003300245851278305, + 0.014413034543395042, + 0.002637176774442196, + 0.005676334258168936, + 0.0037959192413836718, + 0.004301475826650858, + 0.01376472320407629, + 0.018830064684152603, + 0.057610295712947845, + 0.03496987745165825, + 0.02638428844511509, + 0.04434451088309288, + 0.020402831956744194, + 0.031175335869193077, + 0.023240135982632637, + 0.008670274168252945, + 0.008765343576669693, + 0.009002164006233215, + 0.008236897177994251, + 0.0052566020749509335, + 0.006222226191312075, + 0.007858771830797195, + 0.006087804678827524, + 0.009939991869032383, + 0.012027272023260593, + 0.0097209382802248, + 0.010338096879422665, + 0.019555939361453056, + 0.011670971289277077, + 0.01672327145934105, + 0.015367839485406876, + 0.01771608181297779, + 0.01397171150892973, + 0.011191921308636665, + 0.02207372337579727, + 0.018839815631508827, + 0.01277179829776287, + 0.0267159603536129, + 0.013183804228901863, + 0.004425373859703541, + 0.006396317854523659, + 0.006771784741431475, + 0.005367495119571686, + 0.004203701391816139, + 0.0024602245539426804, + 0.004535771440714598, + 0.002075121272355318, + 0.0037259687669575214, + 0.004999551456421614, + 0.004834335297346115, + 0.002900665393099189, + 0.012704250402748585, + 0.0035791932605206966, + 0.003690449520945549, + 0.002675736555829644, + 0.0026052005123347044, + 0.004187467508018017, + 0.003292735433205962, + 0.004545086529105902, + 0.005399747285991907, + 0.0027239928022027016, + 0.00504700792953372, + 0.002778633264824748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0015197168104350567 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010412277188152075 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010412277188152075, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010412277188152075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28604114055633545, + 0.04652250185608864, + 0.07680660486221313, + 0.16113366186618805, + 0.059312425553798676, + 0.051217954605817795, + 0.054528240114450455, + 0.11400492489337921, + 0.05697869509458542, + 0.09345387667417526 + ], + "credit_map": [ + 0.031787797808647156, + 0.026844868436455727, + 0.03537007421255112, + 0.029706662520766258, + 0.04499247297644615, + 0.034453120082616806, + 0.04699692130088806, + 0.035889189690351486, + 0.004441839177161455, + 0.00433692429214716, + 0.00719443941488862, + 0.0048842718824744225, + 0.008040825836360455, + 0.005906801670789719, + 0.006644146982580423, + 0.005073252134025097, + 0.007703143637627363, + 0.004029138013720512, + 0.022134851664304733, + 0.0035912343300879, + 0.004900896456092596, + 0.005593930371105671, + 0.005447382107377052, + 0.023406026884913445, + 0.012659994885325432, + 0.035563576966524124, + 0.020862193778157234, + 0.0205175019800663, + 0.021157754585146904, + 0.013968623243272305, + 0.019346661865711212, + 0.01705733872950077, + 0.005509319249540567, + 0.005313040688633919, + 0.004697027616202831, + 0.005078477784991264, + 0.008601986803114414, + 0.012465759180486202, + 0.00711539713665843, + 0.010531408712267876, + 0.008154050447046757, + 0.0045324512757360935, + 0.004988547880202532, + 0.005960253067314625, + 0.006636389531195164, + 0.005658760201185942, + 0.006266247946768999, + 0.009021247737109661, + 0.006679689045995474, + 0.006617023143917322, + 0.003707562806084752, + 0.005019826348870993, + 0.011000996455550194, + 0.009516265243291855, + 0.006973792798817158, + 0.005013082176446915, + 0.009105202741920948, + 0.011107437312602997, + 0.01832771860063076, + 0.020711343735456467, + 0.012112008407711983, + 0.005640562158077955, + 0.02740374766290188, + 0.00959689635783434, + 0.005146624520421028, + 0.006821595132350922, + 0.008207936771214008, + 0.007746878545731306, + 0.013192723505198956, + 0.0039961738511919975, + 0.006838283501565456, + 0.005028472747653723, + 0.00701143080368638, + 0.010549888014793396, + 0.010830244980752468, + 0.015105465427041054, + 0.01224764995276928, + 0.010005282238125801, + 0.019191615283489227, + 0.008512290194630623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0010412277188152075 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17433783411979675, + 0.12807752192020416, + 0.13345545530319214, + 0.20182117819786072, + 0.033724382519721985, + 0.05146900191903114, + 0.0694081038236618, + 0.050527460873126984, + 0.04811156168580055, + 0.10906752943992615 + ], + "credit_map": [ + 0.01680801995098591, + 0.017949692904949188, + 0.0343552865087986, + 0.013319669291377068, + 0.027309279888868332, + 0.02346481941640377, + 0.020705347880721092, + 0.02042570896446705, + 0.014079490676522255, + 0.008302677422761917, + 0.01400299184024334, + 0.014958728104829788, + 0.02294192463159561, + 0.015672728419303894, + 0.021397368982434273, + 0.016721615567803383, + 0.012236767448484898, + 0.008638492785394192, + 0.03559703752398491, + 0.00806241761893034, + 0.008043735288083553, + 0.007979972288012505, + 0.010586462914943695, + 0.042310573160648346, + 0.017945045605301857, + 0.031542886048555374, + 0.02739318646490574, + 0.031258855015039444, + 0.02886790595948696, + 0.021336540579795837, + 0.025194574147462845, + 0.01828218810260296, + 0.006965660955756903, + 0.004404706880450249, + 0.0033075662795454264, + 0.004275294952094555, + 0.002805299125611782, + 0.004215574823319912, + 0.004275403451174498, + 0.0034748779144138098, + 0.005936719477176666, + 0.003724906127899885, + 0.004738582763820887, + 0.004854908213019371, + 0.01222299225628376, + 0.004666980355978012, + 0.006109157577157021, + 0.00921475701034069, + 0.004813888110220432, + 0.004675433039665222, + 0.0034459379967302084, + 0.012763189151883125, + 0.004786295350641012, + 0.005136856809258461, + 0.027670755982398987, + 0.006115748547017574, + 0.005464488174766302, + 0.009024303406476974, + 0.00719321146607399, + 0.007764515001326799, + 0.00626962399110198, + 0.003011735389009118, + 0.0086086830124259, + 0.0031908999662846327, + 0.0037576123140752316, + 0.005051475949585438, + 0.006423579528927803, + 0.004121432546526194, + 0.0162744652479887, + 0.0046292576007544994, + 0.004566087853163481, + 0.0032876518089324236, + 0.007053301203995943, + 0.018582042306661606, + 0.009874561801552773, + 0.013756713829934597, + 0.022723376750946045, + 0.010466046631336212, + 0.016708875074982643, + 0.009902606718242168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.0009214208112098277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera tilts downward at high speed, captured in extreme slow motion, closely following a crystalline, conical icicle detached from a high rooftop. Encased within tiny bubbles and frozen veins, its mirror-like surface reflects the chilly, clear winter morning light. Plummeting under the weight of gravity, it carves a blurry path through the air, careening towards the dark, frost-covered stone floor below. Upon impact, the icicle's core suddenly explodes, shattering in all directions into countless angular, crystalline fragments of varying sizes. Each fragment, like a miniature diamond, is clearly visible in the slow motion, spinning and tumbling wildly, capturing and refracting the cold morning light. Against the dark background, it creates a fleeting, dazzling haze of light that eventually disperses like stardust. The image is imbued with a pure physical beauty and a chilling, fragile tranquility, dominated by icy blue, dark gray, and high-gloss white.", + "seed": 24003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_6324_seed_24003_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": -0.022156632761470972, + "vlm_logprob": -0.022156632761470972, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.022156632761470972, + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016389507800340652 + ], + "opposite_logprob_tokens": [ + -6.1413893699646 + ], + "logprob_reward": -0.016389507800340652, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.016389507800340652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1070924699306488, + 0.22358354926109314, + 0.0748521164059639, + 0.06569840013980865, + 0.05914032459259033, + 0.15940800309181213, + 0.06730897724628448, + 0.049638763070106506, + 0.053093116730451584, + 0.14018425345420837 + ], + "credit_map": [ + 0.009411431849002838, + 0.009385291486978531, + 0.01461157575249672, + 0.01381982397288084, + 0.015995163470506668, + 0.013073671609163284, + 0.01390690729022026, + 0.01688859611749649, + 0.028888870030641556, + 0.031140821054577827, + 0.026201320812106133, + 0.019245928153395653, + 0.024942943826317787, + 0.028758680447936058, + 0.021416014060378075, + 0.04298898205161095, + 0.0059718419797718525, + 0.011071361601352692, + 0.00867084227502346, + 0.010673106648027897, + 0.009720115922391415, + 0.009747239761054516, + 0.009493629448115826, + 0.009503980167210102, + 0.011195951141417027, + 0.012724969536066055, + 0.009022657759487629, + 0.006435232702642679, + 0.005722732748836279, + 0.006647356785833836, + 0.007170659489929676, + 0.006778836250305176, + 0.007171697448939085, + 0.006923191249370575, + 0.00701733585447073, + 0.009812923148274422, + 0.006047001574188471, + 0.00801462959498167, + 0.005918128415942192, + 0.008235421031713486, + 0.009226428344845772, + 0.011780950240790844, + 0.022664334625005722, + 0.0201321542263031, + 0.007692170329391956, + 0.015543747693300247, + 0.026440201327204704, + 0.04592800885438919, + 0.008408606983721256, + 0.008144715800881386, + 0.00998073723167181, + 0.01040839496999979, + 0.005655390676110983, + 0.008617459796369076, + 0.00784351211041212, + 0.008250164799392223, + 0.0038426655810326338, + 0.0056354934349656105, + 0.0076105473563075066, + 0.010464804247021675, + 0.0034016757272183895, + 0.004623594228178263, + 0.00846864003688097, + 0.005591341760009527, + 0.00413858238607645, + 0.0061048706993460655, + 0.008609910495579243, + 0.009975284337997437, + 0.004335963167250156, + 0.004777122288942337, + 0.009187642484903336, + 0.005963737610727549, + 0.02316882088780403, + 0.021354906260967255, + 0.021088724955916405, + 0.012741701677441597, + 0.007656777277588844, + 0.022431902587413788, + 0.011259732767939568, + 0.020481688901782036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.016389507800340652 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904905617237091 + ], + "opposite_logprob_tokens": [ + -2.4640491008758545 + ], + "logprob_reward": -0.08904905617237091, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904905617237091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07972283661365509, + 0.1544923186302185, + 0.06694359332323074, + 0.3148142397403717, + 0.0702972412109375, + 0.142316997051239, + 0.027898699045181274, + 0.04741717875003815, + 0.02768848091363907, + 0.06840840727090836 + ], + "credit_map": [ + 0.010117396712303162, + 0.006055773701518774, + 0.013702068477869034, + 0.010773630812764168, + 0.0074680340476334095, + 0.009949758648872375, + 0.008981914259493351, + 0.012674262747168541, + 0.019636666402220726, + 0.024894507601857185, + 0.021780023351311684, + 0.011645637452602386, + 0.016296692192554474, + 0.02019333839416504, + 0.01512222457677126, + 0.024923237040638924, + 0.006427749525755644, + 0.012450416572391987, + 0.008129538968205452, + 0.007845772430300713, + 0.00694624288007617, + 0.011259247548878193, + 0.007350775878876448, + 0.006533855572342873, + 0.03720023110508919, + 0.023847414180636406, + 0.065670907497406, + 0.034893013536930084, + 0.02709309197962284, + 0.03221910446882248, + 0.022562844678759575, + 0.07132760435342789, + 0.008102661930024624, + 0.008577099069952965, + 0.007315769791603088, + 0.009004516527056694, + 0.007306605111807585, + 0.013258235529065132, + 0.006748032756149769, + 0.009984317235648632, + 0.0108652887865901, + 0.014355978928506374, + 0.028142880648374557, + 0.021625790745019913, + 0.010661588981747627, + 0.011975521221756935, + 0.01685844175517559, + 0.027831511572003365, + 0.0025052716955542564, + 0.0024226242676377296, + 0.0037562143988907337, + 0.004177159164100885, + 0.002032212447375059, + 0.0028598683420568705, + 0.003626684658229351, + 0.006518663838505745, + 0.0040673459880054, + 0.005257089156657457, + 0.00626810546964407, + 0.008880003355443478, + 0.00414622388780117, + 0.005555152427405119, + 0.007639649324119091, + 0.0056036124005913734, + 0.002183393342420459, + 0.003322359174489975, + 0.004264387302100658, + 0.004700183402746916, + 0.002334370044991374, + 0.003177334787324071, + 0.004211836960166693, + 0.0034946156665682793, + 0.007272572256624699, + 0.009513713419437408, + 0.010942131280899048, + 0.008963007479906082, + 0.006984814070165157, + 0.008277506567537785, + 0.008503690361976624, + 0.007950970903038979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.08904905617237091 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024963426403701305 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.0024963426403701305, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024963426403701305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27268174290657043, + 0.1081201434135437, + 0.08449765294790268, + 0.1387885957956314, + 0.08601591736078262, + 0.08086992800235748, + 0.032660916447639465, + 0.0972796157002449, + 0.02438465692102909, + 0.07470083236694336 + ], + "credit_map": [ + 0.024541709572076797, + 0.015397793613374233, + 0.06695124506950378, + 0.0372128002345562, + 0.020713383331894875, + 0.0319434218108654, + 0.028957515954971313, + 0.046963874250650406, + 0.012074778787791729, + 0.026026854291558266, + 0.01053682528436184, + 0.008241603150963783, + 0.012139566242694855, + 0.01775357685983181, + 0.007054779678583145, + 0.014292154461145401, + 0.007478304207324982, + 0.013996731489896774, + 0.012316578067839146, + 0.01162027008831501, + 0.011103418655693531, + 0.008827335201203823, + 0.009812604635953903, + 0.00934241246432066, + 0.0127570740878582, + 0.008448430337011814, + 0.036921631544828415, + 0.013006647117435932, + 0.009869949892163277, + 0.012737676501274109, + 0.008210018277168274, + 0.03683715686202049, + 0.008562951348721981, + 0.013055156916379929, + 0.00866838451474905, + 0.00930716097354889, + 0.011908712796866894, + 0.01762266457080841, + 0.007237781770527363, + 0.009653103537857533, + 0.005169981624931097, + 0.008217043243348598, + 0.012917493470013142, + 0.011932373046875, + 0.007881667464971542, + 0.006853781640529633, + 0.009863290004432201, + 0.018034303560853004, + 0.002811566460877657, + 0.0036426398437470198, + 0.004100187681615353, + 0.006711343768984079, + 0.0028021864127367735, + 0.0034337444230914116, + 0.004212419502437115, + 0.004946828819811344, + 0.011389825493097305, + 0.010153232142329216, + 0.014024198055267334, + 0.020468013361096382, + 0.00732197193428874, + 0.010387939400970936, + 0.01065109483897686, + 0.012883344665169716, + 0.0017187119228765368, + 0.0023354818113148212, + 0.00386845413595438, + 0.005697529297322035, + 0.0017684469930827618, + 0.002397974720224738, + 0.0027566386852413416, + 0.0038414199370890856, + 0.004327571485191584, + 0.012776828370988369, + 0.025312058627605438, + 0.005640849005430937, + 0.0037240274250507355, + 0.00839877501130104, + 0.008624411188066006, + 0.005896315444260836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0024963426403701305 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015122179174795747 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015122179174795747, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015122179174795747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5263990163803101, + 0.1500730961561203, + 0.04076693579554558, + 0.04414546117186546, + 0.055959757417440414, + 0.05606139823794365, + 0.022006994113326073, + 0.014649922959506512, + 0.013387635350227356, + 0.07654976099729538 + ], + "credit_map": [ + 0.04446258395910263, + 0.030886774882674217, + 0.12975315749645233, + 0.06924092769622803, + 0.041845615953207016, + 0.06410563737154007, + 0.05986446887254715, + 0.08623985946178436, + 0.016984177753329277, + 0.025457419455051422, + 0.018338538706302643, + 0.010355954989790916, + 0.016080956906080246, + 0.020891737192869186, + 0.013853262178599834, + 0.0281110480427742, + 0.003830978646874428, + 0.009111091494560242, + 0.00482794689014554, + 0.004839338827878237, + 0.005100382026284933, + 0.004595944657921791, + 0.003782068146392703, + 0.004679188132286072, + 0.005076131317764521, + 0.008801019750535488, + 0.004404351580888033, + 0.004618247505277395, + 0.005037122871726751, + 0.0054635354317724705, + 0.005510558839887381, + 0.005234491545706987, + 0.006051935721188784, + 0.007464301772415638, + 0.0057320063933730125, + 0.0065685962326824665, + 0.008104057982563972, + 0.008165146224200726, + 0.0061129005625844, + 0.007760814391076565, + 0.0036149928346276283, + 0.004554965533316135, + 0.013664773665368557, + 0.006289868615567684, + 0.003522210754454136, + 0.005590226035565138, + 0.006382325664162636, + 0.01244203932583332, + 0.0015412835637107491, + 0.003074391046538949, + 0.0027354899793863297, + 0.004004388581961393, + 0.0021447923500090837, + 0.00187488307710737, + 0.0031154402531683445, + 0.0035163252614438534, + 0.0012778575764968991, + 0.0014196181437000632, + 0.0025017084553837776, + 0.002548228483647108, + 0.001649667858146131, + 0.0013506243703886867, + 0.0020475969649851322, + 0.001854621572420001, + 0.0011242585023865104, + 0.001168257906101644, + 0.001975565915927291, + 0.002144675701856613, + 0.0010056209284812212, + 0.0018248071428388357, + 0.002287892624735832, + 0.0018565572099760175, + 0.0095934197306633, + 0.005620493553578854, + 0.018035195767879486, + 0.005950190592557192, + 0.006888085510581732, + 0.015606761910021305, + 0.007308794651180506, + 0.0075468141585588455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0015122179174795747 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46450075507164, + 0.1710565984249115, + 0.05156943202018738, + 0.040843185037374496, + 0.037719130516052246, + 0.10415790230035782, + 0.04370708391070366, + 0.03293352201581001, + 0.015714121982455254, + 0.037798307836055756 + ], + "credit_map": [ + 0.038860321044921875, + 0.02652040496468544, + 0.11456038802862167, + 0.06143607199192047, + 0.03819369897246361, + 0.05785106122493744, + 0.05250097066164017, + 0.07457783818244934, + 0.021012548357248306, + 0.03643403202295303, + 0.01689095050096512, + 0.01390470564365387, + 0.017764614894986153, + 0.02539856731891632, + 0.01326770056039095, + 0.026383478194475174, + 0.004413269925862551, + 0.006225523538887501, + 0.007697715889662504, + 0.007411622907966375, + 0.0069623966701328754, + 0.006340535823255777, + 0.0064135221764445305, + 0.0061048404313623905, + 0.005465262103825808, + 0.005048207473009825, + 0.00833093747496605, + 0.0028599274810403585, + 0.0038855422753840685, + 0.004665067419409752, + 0.002929974114522338, + 0.007658260874450207, + 0.005566893611103296, + 0.005464090500026941, + 0.0033607915975153446, + 0.004254171624779701, + 0.003940444905310869, + 0.00534004345536232, + 0.004283788148313761, + 0.005508902482688427, + 0.007040269672870636, + 0.007649488281458616, + 0.015456424094736576, + 0.01340172067284584, + 0.004864408168941736, + 0.008426455780863762, + 0.0148488599807024, + 0.03247026726603508, + 0.004147435538470745, + 0.0037229941226541996, + 0.00925162062048912, + 0.006342091131955385, + 0.003823030972853303, + 0.0049398778937757015, + 0.005838859360665083, + 0.005641171708703041, + 0.0025991082657128572, + 0.002909767907112837, + 0.00469239940866828, + 0.006626071874052286, + 0.0023529899772256613, + 0.0030616638250648975, + 0.004879984073340893, + 0.005811535753309727, + 0.000978260999545455, + 0.0012166276574134827, + 0.0028168908320367336, + 0.0024608634412288666, + 0.0015812369529157877, + 0.001466538175009191, + 0.0031336310785263777, + 0.0020600720308721066, + 0.003405192866921425, + 0.003506086068227887, + 0.013616244308650494, + 0.003929616883397102, + 0.0023288128431886435, + 0.0026039956137537956, + 0.005836577620357275, + 0.00257178139872849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0013360392767935991 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera tilts downward at high speed, captured in extreme slow motion, closely following a crystalline, conical icicle detached from a high rooftop. Encased within tiny bubbles and frozen veins, its mirror-like surface reflects the chilly, clear winter morning light. Plummeting under the weight of gravity, it carves a blurry path through the air, careening towards the dark, frost-covered stone floor below. Upon impact, the icicle's core suddenly explodes, shattering in all directions into countless angular, crystalline fragments of varying sizes. Each fragment, like a miniature diamond, is clearly visible in the slow motion, spinning and tumbling wildly, capturing and refracting the cold morning light. Against the dark background, it creates a fleeting, dazzling haze of light that eventually disperses like stardust. The image is imbued with a pure physical beauty and a chilling, fragile tranquility, dominated by icy blue, dark gray, and high-gloss white.", + "seed": 24004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_12/index_6324_seed_24004_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": -0.08247764692641794, + "vlm_logprob": -0.08247764692641794, + "vlm_yes_no_margin": 4.774999988079071, + "vlm_token_credit": -0.08247764692641794, + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016531866043806076 + ], + "opposite_logprob_tokens": [ + -5.016531944274902 + ], + "logprob_reward": -0.016531866043806076, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016531866043806076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214386597275734, + 0.05833296850323677, + 0.09122121334075928, + 0.15258906781673431, + 0.06433912366628647, + 0.08232775330543518, + 0.19037936627864838, + 0.0787874385714531, + 0.036322616040706635, + 0.12426178902387619 + ], + "credit_map": [ + 0.018982475623488426, + 0.015741031616926193, + 0.009724683128297329, + 0.012094803154468536, + 0.01598718762397766, + 0.015192950144410133, + 0.01320117898285389, + 0.020514344796538353, + 0.005062463227659464, + 0.00705516804009676, + 0.004153682384639978, + 0.00932103767991066, + 0.008547068573534489, + 0.007294940762221813, + 0.010498845018446445, + 0.006399763282388449, + 0.010793862864375114, + 0.011746706441044807, + 0.011530105955898762, + 0.012289485894143581, + 0.01051393337547779, + 0.011128164827823639, + 0.014403867535293102, + 0.00881508644670248, + 0.02340732142329216, + 0.009615383110940456, + 0.028572509065270424, + 0.010507979430258274, + 0.01505989395081997, + 0.014598431065678596, + 0.03219304978847504, + 0.018634505569934845, + 0.006852033548057079, + 0.016211552545428276, + 0.006792963948100805, + 0.006665782071650028, + 0.00907820463180542, + 0.00666479766368866, + 0.007989258505403996, + 0.00408452283591032, + 0.007099454291164875, + 0.0065348283387720585, + 0.013259273022413254, + 0.011206213384866714, + 0.010083268396556377, + 0.01419871486723423, + 0.011214861646294594, + 0.008731142617762089, + 0.01132028829306364, + 0.05859221890568733, + 0.07010252773761749, + 0.01458122581243515, + 0.007847435772418976, + 0.009595000185072422, + 0.010538175702095032, + 0.0078024836257100105, + 0.009340082295238972, + 0.020268281921744347, + 0.011549093760550022, + 0.007780694402754307, + 0.006694633048027754, + 0.008693091571331024, + 0.005371821112930775, + 0.009089729748666286, + 0.007357319816946983, + 0.0051626116037368774, + 0.0034164567478001118, + 0.0036742519587278366, + 0.004183103796094656, + 0.004038848914206028, + 0.0030763891991227865, + 0.005413633771240711, + 0.010912369936704636, + 0.013321677222847939, + 0.020363664254546165, + 0.0171239972114563, + 0.015611011534929276, + 0.009993601590394974, + 0.011608555912971497, + 0.025326915085315704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.016531866043806076 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38706547021865845 + ], + "opposite_logprob_tokens": [ + -1.1370654106140137 + ], + "logprob_reward": -0.38706547021865845, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38706547021865845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12749601900577545, + 0.13109830021858215, + 0.1076648160815239, + 0.10268007218837738, + 0.05736565962433815, + 0.11065752059221268, + 0.1939181238412857, + 0.05856897309422493, + 0.066987544298172, + 0.043563004583120346 + ], + "credit_map": [ + 0.022013792768120766, + 0.015745684504508972, + 0.01195396576076746, + 0.012201801873743534, + 0.015355831943452358, + 0.015702808275818825, + 0.011912270449101925, + 0.022609859704971313, + 0.01281545776873827, + 0.015994612127542496, + 0.010275231674313545, + 0.019635364413261414, + 0.015307756140828133, + 0.016969231888651848, + 0.02523023635149002, + 0.014870412647724152, + 0.01198471151292324, + 0.009642444550991058, + 0.0267975851893425, + 0.008010867051780224, + 0.006876329891383648, + 0.013445157557725906, + 0.012736724689602852, + 0.018170993775129318, + 0.01416053343564272, + 0.0071684070862829685, + 0.02437697723507881, + 0.008349827490746975, + 0.00923070777207613, + 0.00876552052795887, + 0.018008748069405556, + 0.012619352899491787, + 0.005843799561262131, + 0.00985574722290039, + 0.0073712109588086605, + 0.006960084196180105, + 0.006372543051838875, + 0.00934704951941967, + 0.0058110561221838, + 0.0058041648007929325, + 0.008803953416645527, + 0.010268022306263447, + 0.012189291417598724, + 0.01781851425766945, + 0.011389657855033875, + 0.023694733157753944, + 0.01222815178334713, + 0.014265195466578007, + 0.01174912042915821, + 0.055988896638154984, + 0.07856644690036774, + 0.013725711964070797, + 0.006989297457039356, + 0.008899510838091373, + 0.010967935435473919, + 0.007031215820461512, + 0.007392253261059523, + 0.01145282294601202, + 0.010285337455570698, + 0.009523172862827778, + 0.003992724232375622, + 0.005222799256443977, + 0.004034813493490219, + 0.006665045395493507, + 0.016449986025691032, + 0.008697346784174442, + 0.007140332832932472, + 0.005420201923698187, + 0.011944160796701908, + 0.00756890419870615, + 0.00533251790329814, + 0.0044340938329696655, + 0.004906844813376665, + 0.006281409412622452, + 0.005255117546766996, + 0.003493885276839137, + 0.005506767425686121, + 0.00530250696465373, + 0.004430663771927357, + 0.00838580820709467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.38706547021865845 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002495629247277975 + ], + "opposite_logprob_tokens": [ + -6.002495765686035 + ], + "logprob_reward": -0.002495629247277975, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002495629247277975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11179623752832413, + 0.09427797794342041, + 0.05629235878586769, + 0.08680431544780731, + 0.20331282913684845, + 0.08047151565551758, + 0.11392603814601898, + 0.08862614631652832, + 0.10557610541582108, + 0.058916475623846054 + ], + "credit_map": [ + 0.013594320975244045, + 0.013446830213069916, + 0.010578503832221031, + 0.008674521930515766, + 0.020454296842217445, + 0.014886484481394291, + 0.01075157430022955, + 0.019409697502851486, + 0.013151776045560837, + 0.008116672746837139, + 0.008703397586941719, + 0.011209368705749512, + 0.008211501874029636, + 0.008871088735759258, + 0.014013681560754776, + 0.022000489756464958, + 0.006110706832259893, + 0.0056649004109203815, + 0.010272976011037827, + 0.006012580823153257, + 0.00418646028265357, + 0.007406212855130434, + 0.007931418716907501, + 0.008707106113433838, + 0.013281198218464851, + 0.0072709652595222, + 0.009907630272209644, + 0.007301679812371731, + 0.01026748027652502, + 0.009654740802943707, + 0.018903639167547226, + 0.010216975584626198, + 0.019724631682038307, + 0.03669260814785957, + 0.02432256005704403, + 0.03220163658261299, + 0.021056020632386208, + 0.02628181129693985, + 0.02749142423272133, + 0.015542126260697842, + 0.009263609535992146, + 0.01124771311879158, + 0.009914073161780834, + 0.011772319674491882, + 0.009483990259468555, + 0.0123981898650527, + 0.008638007566332817, + 0.0077536148019135, + 0.007595825009047985, + 0.015042832121253014, + 0.02993503026664257, + 0.012125959619879723, + 0.012442552484571934, + 0.009566963650286198, + 0.009138857945799828, + 0.01807801052927971, + 0.01599070429801941, + 0.012910786084830761, + 0.015771858394145966, + 0.017444096505641937, + 0.006857366766780615, + 0.007011230103671551, + 0.005514195654541254, + 0.00712590804323554, + 0.015662454068660736, + 0.013084582053124905, + 0.014319585636258125, + 0.011284801177680492, + 0.012963024899363518, + 0.016174791380763054, + 0.010794218629598618, + 0.011292633600533009, + 0.006230788305401802, + 0.008851513266563416, + 0.008236034773290157, + 0.006781489122658968, + 0.009016327559947968, + 0.005422619171440601, + 0.006546202581375837, + 0.007831498980522156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.002495629247277975 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251778755337 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251778755337, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251778755337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0441766232252121, + 0.03489883616566658, + 0.09784676879644394, + 0.1303752064704895, + 0.15296325087547302, + 0.07626806199550629, + 0.08304403722286224, + 0.11209113150835037, + 0.08165691792964935, + 0.1866791844367981 + ], + "credit_map": [ + 0.005770354997366667, + 0.005283370614051819, + 0.004308162722736597, + 0.004076195880770683, + 0.00719882408156991, + 0.0050291381776332855, + 0.0050931526347994804, + 0.007417420856654644, + 0.00473899208009243, + 0.0027629933319985867, + 0.0027416865341365337, + 0.003309632884338498, + 0.0034801852889358997, + 0.00344484718516469, + 0.007137729320675135, + 0.007282767910510302, + 0.015011929906904697, + 0.01035468839108944, + 0.01323406770825386, + 0.007667813450098038, + 0.01360659021884203, + 0.01526977401226759, + 0.012119133025407791, + 0.01058276928961277, + 0.020907606929540634, + 0.006270081736147404, + 0.019516626372933388, + 0.009140649810433388, + 0.013131479732692242, + 0.013063188642263412, + 0.034919071942567825, + 0.013426501303911209, + 0.014859785325825214, + 0.028047220781445503, + 0.022345801815390587, + 0.01723361574113369, + 0.016381653025746346, + 0.028120409697294235, + 0.016771206632256508, + 0.009203546680510044, + 0.005006037652492523, + 0.007400680333375931, + 0.014269760809838772, + 0.010210899636149406, + 0.008519677445292473, + 0.013984021730720997, + 0.008192365057766438, + 0.008684622123837471, + 0.009790456853806973, + 0.013543051667511463, + 0.012885097414255142, + 0.009561162441968918, + 0.011172340251505375, + 0.009788679890334606, + 0.007268517278134823, + 0.009034732356667519, + 0.011356711387634277, + 0.029994171112775803, + 0.030909331515431404, + 0.01798451691865921, + 0.006642731372267008, + 0.005036991089582443, + 0.005546648986637592, + 0.004620037507265806, + 0.013389305211603642, + 0.009253480471670628, + 0.010989449918270111, + 0.009132078848779202, + 0.008272980339825153, + 0.013686459511518478, + 0.009352799504995346, + 0.0075803580693900585, + 0.012609428726136684, + 0.023835288360714912, + 0.02142505906522274, + 0.036085814237594604, + 0.0215641837567091, + 0.017089881002902985, + 0.017840705811977386, + 0.03622882813215256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.005251778755337 + }, + { + "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 crystalline, conical icicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera tilting downward at high speed while capturing the scene in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the icicle have a mirror-like surface with visible tiny bubbles and frozen veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the icicle shatter into countless angular, crystalline fragments upon impact with the dark, frost-covered stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by icy blue, dark gray, and high-gloss white colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010434903670102358 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010434903670102358, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010434903670102358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12047844380140305, + 0.10305579751729965, + 0.047899279743433, + 0.15014195442199707, + 0.058764275163412094, + 0.0442379005253315, + 0.15784886479377747, + 0.08947869390249252, + 0.07661008834838867, + 0.15148471295833588 + ], + "credit_map": [ + 0.012533724308013916, + 0.018621310591697693, + 0.010374387726187706, + 0.014810148626565933, + 0.01746022142469883, + 0.015225880779325962, + 0.015631122514605522, + 0.015821650624275208, + 0.02075078710913658, + 0.006864108610898256, + 0.010450595058500767, + 0.0063641853630542755, + 0.006656529847532511, + 0.009235507808625698, + 0.010987362824380398, + 0.03174671530723572, + 0.0053704106248915195, + 0.004984027240425348, + 0.009603829123079777, + 0.0038586666341871023, + 0.004063972271978855, + 0.00708443857729435, + 0.0054375045001506805, + 0.007496435195207596, + 0.02496570162475109, + 0.008031832054257393, + 0.023373063653707504, + 0.009529117494821548, + 0.014974366873502731, + 0.012502693571150303, + 0.03041135147213936, + 0.02635382115840912, + 0.005896424874663353, + 0.0056019313633441925, + 0.010303465649485588, + 0.008038253523409367, + 0.00779726030305028, + 0.01107380073517561, + 0.006246872711926699, + 0.003806266002357006, + 0.0042580231092870235, + 0.006185926962643862, + 0.0073414333164691925, + 0.006581312045454979, + 0.00585797056555748, + 0.005653835833072662, + 0.004302769433706999, + 0.004056631587445736, + 0.011368809267878532, + 0.035909365862607956, + 0.05392170324921608, + 0.013563345186412334, + 0.011418314650654793, + 0.011080974712967873, + 0.010096791200339794, + 0.010489562526345253, + 0.010792658664286137, + 0.017476463690400124, + 0.019288670271635056, + 0.013653573580086231, + 0.006997978314757347, + 0.006525607779622078, + 0.005312981549650431, + 0.009430760517716408, + 0.011045720428228378, + 0.006632356438785791, + 0.009931429289281368, + 0.007299148943275213, + 0.014853422529995441, + 0.013183575123548508, + 0.006601785309612751, + 0.007062647026032209, + 0.014518847689032555, + 0.015947168692946434, + 0.028429130092263222, + 0.01219948474317789, + 0.017559295520186424, + 0.011734016239643097, + 0.018452145159244537, + 0.03264462947845459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.0010434903670102358 + } + ], + "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/0000060/group_13/index_tensor([9776, 9776])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_tensor([9776, 9776])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f4c83d1dbd8b443a361d3b0c51c965ec74d5be6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_tensor([9776, 9776])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a panoramic shot rendered in a polished Pixar style, the camera glides smoothly in an arc around a central, highly polished, dark wooden stage. At its center, the joyful taiko character, Wadadou, performs a playful dance. His body, a perfectly spherical light brown drum with subtle textures, gleams under the bright, warm stage lights. A dazzling cartoon smile perpetually etched on his face, flanked by two round, rosy cheeks, while his large, simple black eyes sparkle with pure joy. He bounces rhythmically on his base, his entire round body swaying slightly with each jump, while his short, slender, cylindrical limbs pulsate with exuberant, carefree energy. The spotlessly polished stage floor reflects his every movement in a soft, vibrant color. Against the vibrant, soft-focus background, colorful hanging banners and glowing paper lanterns fill the festive world, creating a scene awash in saturated, warm colors. The entire image is bathed in a soft, celebratory light that eliminates harsh shadows, creating an overwhelmingly joyous and energetic atmosphere of pure, infectious happiness.", + "seed": 28284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28284_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": -0.02276662169024348, + "vlm_logprob": -0.02276662169024348, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02276662169024348, + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0358349084854126 + ], + "opposite_logprob_tokens": [ + -3.785834789276123 + ], + "logprob_reward": -0.0358349084854126, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0358349084854126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0732395201921463, + 0.09923751652240753, + 0.04153601825237274, + 0.07690557837486267, + 0.19250048696994781, + 0.17023199796676636, + 0.12254917621612549, + 0.09842333942651749, + 0.0682201161980629, + 0.05715617164969444 + ], + "credit_map": [ + 0.009019821882247925, + 0.007646598853170872, + 0.00572106521576643, + 0.008058134466409683, + 0.013535050675272942, + 0.010608838871121407, + 0.008304032497107983, + 0.010345982387661934, + 0.015988513827323914, + 0.012354107573628426, + 0.0075535555370152, + 0.013451876118779182, + 0.013517243787646294, + 0.012285297736525536, + 0.014747122302651405, + 0.009339800104498863, + 0.00849977508187294, + 0.004779146984219551, + 0.00446936022490263, + 0.005422916729003191, + 0.003546124789863825, + 0.004710955545306206, + 0.005951478146016598, + 0.0041562593542039394, + 0.007674825843423605, + 0.010342450812458992, + 0.009530276991426945, + 0.008050474338233471, + 0.011733551509678364, + 0.011192597448825836, + 0.010560858994722366, + 0.00782053917646408, + 0.056625936180353165, + 0.006128434091806412, + 0.022007448598742485, + 0.007401754148304462, + 0.004548406694084406, + 0.016535023227334023, + 0.009498934261500835, + 0.06975456327199936, + 0.01635911874473095, + 0.019916726276278496, + 0.015264133922755718, + 0.021737294271588326, + 0.016649406403303146, + 0.04114184156060219, + 0.017052730545401573, + 0.022110743448138237, + 0.010644269175827503, + 0.009938199073076248, + 0.007566193118691444, + 0.009402787312865257, + 0.009081894531846046, + 0.02838600054383278, + 0.012813174165785313, + 0.034716665744781494, + 0.006305862218141556, + 0.007405891083180904, + 0.028881004080176353, + 0.007812692783772945, + 0.00688942801207304, + 0.0074926414526999, + 0.0066036563366651535, + 0.027032168582081795, + 0.010739784687757492, + 0.008777469396591187, + 0.00733587983995676, + 0.006334607023745775, + 0.008112474344670773, + 0.010954417288303375, + 0.008138925768435001, + 0.007826559245586395, + 0.0077490913681685925, + 0.004682521335780621, + 0.006503204815089703, + 0.005906886421144009, + 0.012549081817269325, + 0.005668996833264828, + 0.00709868548437953, + 0.006997707299888134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0358349084854126 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0705459862947464, + 0.18823595345020294, + 0.027084002271294594, + 0.040533918887376785, + 0.1283382922410965, + 0.06467688828706741, + 0.3612673282623291, + 0.03410227596759796, + 0.036782022565603256, + 0.04843340441584587 + ], + "credit_map": [ + 0.007183007895946503, + 0.009059232659637928, + 0.005427802912890911, + 0.008460471406579018, + 0.012517751194536686, + 0.009695383720099926, + 0.009334714151918888, + 0.008867619559168816, + 0.02415110543370247, + 0.0253983736038208, + 0.011723875068128109, + 0.023742420598864555, + 0.03346864506602287, + 0.028602926060557365, + 0.028130022808909416, + 0.013018582947552204, + 0.004327526316046715, + 0.0042035263031721115, + 0.002908738562837243, + 0.0032870755530893803, + 0.0034237641375511885, + 0.002338134218007326, + 0.004069795366376638, + 0.00252544111572206, + 0.004217402543872595, + 0.0049322620034217834, + 0.005007683765143156, + 0.00480789877474308, + 0.005862614139914513, + 0.005074887536466122, + 0.004150652792304754, + 0.006480516865849495, + 0.03779304400086403, + 0.0046895998530089855, + 0.01304685790091753, + 0.005890330765396357, + 0.003772332798689604, + 0.013159445486962795, + 0.006361715029925108, + 0.04362497106194496, + 0.0066876159980893135, + 0.0068503208458423615, + 0.007249743212014437, + 0.011225646361708641, + 0.009606335312128067, + 0.007517124526202679, + 0.008458676747977734, + 0.007081430871039629, + 0.03582637384533882, + 0.042337290942668915, + 0.025456611067056656, + 0.05113256722688675, + 0.0406167134642601, + 0.05068899318575859, + 0.05123424157500267, + 0.0639745220541954, + 0.003255728865042329, + 0.002635055920109153, + 0.007811852730810642, + 0.0027647337410598993, + 0.0033114126417785883, + 0.003032410517334938, + 0.0025306870229542255, + 0.008760396391153336, + 0.005606865044683218, + 0.003848151070997119, + 0.002323953900486231, + 0.0038397919852286577, + 0.007078864611685276, + 0.005413914564996958, + 0.003926943521946669, + 0.004743536934256554, + 0.006175640504807234, + 0.005662834737449884, + 0.0054283058270812035, + 0.0056932661682367325, + 0.011026587337255478, + 0.0033805714920163155, + 0.005891088861972094, + 0.005175111815333366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.011083701625466347 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160687193274498 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160687193274498, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160687193274498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051189299672842026, + 0.04924341291189194, + 0.04252715781331062, + 0.0837550088763237, + 0.3707481920719147, + 0.06340359151363373, + 0.11603915691375732, + 0.06872148811817169, + 0.1221974566578865, + 0.03217525780200958 + ], + "credit_map": [ + 0.0078053465113043785, + 0.006457820534706116, + 0.005091411992907524, + 0.0068029360845685005, + 0.007299016695469618, + 0.005810599774122238, + 0.006355839781463146, + 0.005566329695284367, + 0.005530927795916796, + 0.008145922794938087, + 0.0037496420554816723, + 0.005771676544100046, + 0.009414774365723133, + 0.005821805913001299, + 0.006770249921828508, + 0.004038418177515268, + 0.007413656450808048, + 0.004037429112941027, + 0.004813899751752615, + 0.006524004507809877, + 0.003878759453073144, + 0.004819327034056187, + 0.0068877809680998325, + 0.004152299836277962, + 0.01896650530397892, + 0.005708684213459492, + 0.02692767232656479, + 0.005818170960992575, + 0.006713255308568478, + 0.00722796143963933, + 0.005633082240819931, + 0.006759684532880783, + 0.12235786765813828, + 0.012303355149924755, + 0.032211784273386, + 0.011405806988477707, + 0.008888501673936844, + 0.05341646447777748, + 0.016351494938135147, + 0.11381292343139648, + 0.006829171907156706, + 0.005495710298418999, + 0.006129886023700237, + 0.005660806782543659, + 0.008240777999162674, + 0.0161705631762743, + 0.005568607710301876, + 0.00930806528776884, + 0.010484360158443451, + 0.009037704207003117, + 0.007333575747907162, + 0.012535142712295055, + 0.013215063139796257, + 0.02851029857993126, + 0.012559197843074799, + 0.022363806143403053, + 0.004190061707049608, + 0.008881214074790478, + 0.00691630132496357, + 0.007734971586614847, + 0.009371220134198666, + 0.011220601387321949, + 0.006450984161347151, + 0.013956126756966114, + 0.03419017419219017, + 0.007619106210768223, + 0.007530156522989273, + 0.008604086935520172, + 0.011653833091259003, + 0.012840966694056988, + 0.01139013096690178, + 0.02836899645626545, + 0.004469095263630152, + 0.003313979599624872, + 0.00586690055206418, + 0.003819623962044716, + 0.004940384533256292, + 0.0019772422965615988, + 0.00429943110793829, + 0.003488599555566907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.0160687193274498 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865080490708351 + ], + "opposite_logprob_tokens": [ + -3.0486507415771484 + ], + "logprob_reward": -0.04865080490708351, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865080490708351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287260353565216, + 0.13665291666984558, + 0.11609669774770737, + 0.0558018833398819, + 0.10243529081344604, + 0.09800131618976593, + 0.05007260665297508, + 0.0953742116689682, + 0.14262087643146515, + 0.07421819120645523 + ], + "credit_map": [ + 0.023091580718755722, + 0.010168075561523438, + 0.02420489676296711, + 0.0112753976136446, + 0.014996328391134739, + 0.01621796377003193, + 0.01330627128481865, + 0.015465510077774525, + 0.02102694846689701, + 0.014902712777256966, + 0.011097925715148449, + 0.02090788446366787, + 0.013286586850881577, + 0.021480394527316093, + 0.01828787289559841, + 0.01566259190440178, + 0.0219376552850008, + 0.01519035454839468, + 0.018026551231741905, + 0.014582047238945961, + 0.010431122966110706, + 0.01011247094720602, + 0.011951790191233158, + 0.013864699751138687, + 0.0069877831265330315, + 0.006384523585438728, + 0.007423166185617447, + 0.007284460589289665, + 0.0039275395683944225, + 0.011005950160324574, + 0.0059873200953006744, + 0.0068011414259672165, + 0.009084404446184635, + 0.0040167951956391335, + 0.014398572966456413, + 0.007471087388694286, + 0.004765361547470093, + 0.023643851280212402, + 0.007895003072917461, + 0.031160207465291023, + 0.013771554455161095, + 0.006380208767950535, + 0.013461335562169552, + 0.00996443536132574, + 0.00801230687648058, + 0.01803634501993656, + 0.009572249837219715, + 0.018802877515554428, + 0.0069517409428954124, + 0.0064211334101855755, + 0.005814640782773495, + 0.0071729691699147224, + 0.005337323062121868, + 0.008081172592937946, + 0.005476071964949369, + 0.004817554727196693, + 0.017746180295944214, + 0.008771047927439213, + 0.016546273604035378, + 0.009178613312542439, + 0.012158576399087906, + 0.012386787682771683, + 0.011406619101762772, + 0.0071801175363361835, + 0.015472307801246643, + 0.017205018550157547, + 0.013878744095563889, + 0.01513089518994093, + 0.02778695896267891, + 0.021232591941952705, + 0.018255095928907394, + 0.013659262098371983, + 0.013353344053030014, + 0.005897399969398975, + 0.010733189061284065, + 0.0074752578511834145, + 0.007525011897087097, + 0.008990829810500145, + 0.008245966397225857, + 0.011997192166745663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04865080490708351 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021949741058051586 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021949741058051586, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021949741058051586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06647566705942154, + 0.06417492777109146, + 0.03057185187935829, + 0.09757862985134125, + 0.22433750331401825, + 0.052937909960746765, + 0.05098191648721695, + 0.1165887713432312, + 0.2372513711452484, + 0.05910148099064827 + ], + "credit_map": [ + 0.01220177672803402, + 0.004526243079453707, + 0.005724309012293816, + 0.005003547295928001, + 0.006886234972625971, + 0.013296605087816715, + 0.005339726340025663, + 0.01349722035229206, + 0.011561560444533825, + 0.005720682442188263, + 0.004881908651441336, + 0.009780295193195343, + 0.005325560923665762, + 0.014376448467373848, + 0.0073442221619188786, + 0.005184253212064505, + 0.005286552011966705, + 0.004278091713786125, + 0.003944649826735258, + 0.0037022626493126154, + 0.0029592986684292555, + 0.002988871419802308, + 0.004179026931524277, + 0.00323309819214046, + 0.0100774010643363, + 0.008818223141133785, + 0.015167818404734135, + 0.007809814531356096, + 0.009246044792234898, + 0.020861884579062462, + 0.011411960236728191, + 0.014185487292706966, + 0.0596010684967041, + 0.006456585600972176, + 0.029010998085141182, + 0.00806514173746109, + 0.0063927858136594296, + 0.014584518037736416, + 0.011052606627345085, + 0.08917379379272461, + 0.004196952097117901, + 0.003466659924015403, + 0.004368973895907402, + 0.004684803541749716, + 0.003474706318229437, + 0.0159082543104887, + 0.004468826111406088, + 0.012368732132017612, + 0.0049824840389192104, + 0.009161196649074554, + 0.005926263984292746, + 0.006636978592723608, + 0.006293027196079493, + 0.006617492064833641, + 0.005779200233519077, + 0.005585274659097195, + 0.01316525787115097, + 0.006099645514041185, + 0.033496011048555374, + 0.011097273789346218, + 0.006660165265202522, + 0.013825040310621262, + 0.007225149776786566, + 0.025020232424139977, + 0.038368597626686096, + 0.025090277194976807, + 0.016781214624643326, + 0.022078512236475945, + 0.04364578053355217, + 0.029870036989450455, + 0.026325881481170654, + 0.035091061145067215, + 0.01103256642818451, + 0.00402333727106452, + 0.008174204267561436, + 0.00567319430410862, + 0.006104828789830208, + 0.008666036650538445, + 0.008663720451295376, + 0.006763591431081295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.0021949741058051586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a panoramic shot rendered in a polished Pixar style, the camera glides smoothly in an arc around a central, highly polished, dark wooden stage. At its center, the joyful taiko character, Wadadou, performs a playful dance. His body, a perfectly spherical light brown drum with subtle textures, gleams under the bright, warm stage lights. A dazzling cartoon smile perpetually etched on his face, flanked by two round, rosy cheeks, while his large, simple black eyes sparkle with pure joy. He bounces rhythmically on his base, his entire round body swaying slightly with each jump, while his short, slender, cylindrical limbs pulsate with exuberant, carefree energy. The spotlessly polished stage floor reflects his every movement in a soft, vibrant color. Against the vibrant, soft-focus background, colorful hanging banners and glowing paper lanterns fill the festive world, creating a scene awash in saturated, warm colors. The entire image is bathed in a soft, celebratory light that eliminates harsh shadows, creating an overwhelmingly joyous and energetic atmosphere of pure, infectious happiness.", + "seed": 28285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28285_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": -0.0593314622528851, + "vlm_logprob": -0.0593314622528851, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0593314622528851, + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09293968230485916 + ], + "opposite_logprob_tokens": [ + -2.592939615249634 + ], + "logprob_reward": -0.09293968230485916, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09293968230485916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18790796399116516, + 0.0495796836912632, + 0.0524306446313858, + 0.19916871190071106, + 0.25231432914733887, + 0.09932337701320648, + 0.057024821639060974, + 0.026243869215250015, + 0.05353035032749176, + 0.022476281970739365 + ], + "credit_map": [ + 0.009826956316828728, + 0.010680710896849632, + 0.038943834602832794, + 0.029077941551804543, + 0.012305363081395626, + 0.024357391521334648, + 0.017333194613456726, + 0.0453825481235981, + 0.006442573852837086, + 0.004586982540786266, + 0.005029371473938227, + 0.007434688042849302, + 0.0066405669786036015, + 0.00789631623774767, + 0.006856013089418411, + 0.004693167749792337, + 0.006832551211118698, + 0.008515697903931141, + 0.007687961217015982, + 0.006337687838822603, + 0.004452931694686413, + 0.0048506292514503, + 0.007465734146535397, + 0.006287445314228535, + 0.012006406672298908, + 0.008733689785003662, + 0.016302598640322685, + 0.021339051425457, + 0.016004303470253944, + 0.011108417995274067, + 0.07740765064954758, + 0.03626658767461777, + 0.014618094079196453, + 0.02969568781554699, + 0.008856495842337608, + 0.013631465844810009, + 0.08123069256544113, + 0.07286392897367477, + 0.016636008396744728, + 0.014781955629587173, + 0.00738907465711236, + 0.007895533926784992, + 0.005521412007510662, + 0.03325214982032776, + 0.00975963193923235, + 0.005291249603033066, + 0.01972813718020916, + 0.010486180894076824, + 0.004755495581775904, + 0.009067568928003311, + 0.004700220189988613, + 0.007728923112154007, + 0.008099744096398354, + 0.007853643037378788, + 0.008513222448527813, + 0.00630599819123745, + 0.003127817064523697, + 0.002770965453237295, + 0.004022781737148762, + 0.0029045077972114086, + 0.0036001629196107388, + 0.0028495141305029392, + 0.0035392153076827526, + 0.0034289020113646984, + 0.005012734793126583, + 0.004164539277553558, + 0.007903280667960644, + 0.005021701101213694, + 0.013530265539884567, + 0.005316287279129028, + 0.007978915236890316, + 0.004602622706443071, + 0.0025997376069426537, + 0.002204493386670947, + 0.0033321732189506292, + 0.0035030199214816093, + 0.002653185511007905, + 0.0021182734053581953, + 0.0040882546454668045, + 0.0019771421793848276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.09293968230485916 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109242532402277 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.01109242532402277, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109242532402277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21333451569080353, + 0.030102664604783058, + 0.021529382094740868, + 0.1217896044254303, + 0.42864951491355896, + 0.09897779673337936, + 0.010714812204241753, + 0.019457126036286354, + 0.026945462450385094, + 0.028499111533164978 + ], + "credit_map": [ + 0.013497697189450264, + 0.017600886523723602, + 0.044253118336200714, + 0.03308277204632759, + 0.01415552943944931, + 0.026000363752245903, + 0.018880462273955345, + 0.045863717794418335, + 0.00300455279648304, + 0.004901329055428505, + 0.0028670814353972673, + 0.004230407997965813, + 0.0058336928486824036, + 0.0029957997612655163, + 0.0029452331364154816, + 0.003324570832774043, + 0.0024846941232681274, + 0.0036245090886950493, + 0.0027203455101698637, + 0.0029934789054095745, + 0.002754979534074664, + 0.001911583705805242, + 0.00307742552831769, + 0.0019623683765530586, + 0.007275215350091457, + 0.008844657801091671, + 0.008195198141038418, + 0.012268059886991978, + 0.012131786905229092, + 0.006422373466193676, + 0.03915112838149071, + 0.0275011844933033, + 0.022409116849303246, + 0.04783543944358826, + 0.013552896678447723, + 0.024295810610055923, + 0.14205394685268402, + 0.12577849626541138, + 0.028776712715625763, + 0.023947108536958694, + 0.006114943418651819, + 0.008596781641244888, + 0.006621727719902992, + 0.0269122626632452, + 0.015329350717365742, + 0.007632214576005936, + 0.020863564684987068, + 0.006906956899911165, + 0.000917924044188112, + 0.0012737383367493749, + 0.0010016887681558728, + 0.0014065445866435766, + 0.0011386999394744635, + 0.001327883917838335, + 0.0018354260828346014, + 0.0018129067029803991, + 0.0016686383169144392, + 0.002524725627154112, + 0.0022132620215415955, + 0.0023144222795963287, + 0.003418033942580223, + 0.0021413620561361313, + 0.003090319922193885, + 0.0020863644313067198, + 0.004003475420176983, + 0.0027099407743662596, + 0.0057652839459478855, + 0.002428640378639102, + 0.003714684396982193, + 0.0023774912115186453, + 0.003283063182607293, + 0.002662885468453169, + 0.0033750119619071484, + 0.0030393085908144712, + 0.004061104264110327, + 0.003484490793198347, + 0.00477294996380806, + 0.0026472790632396936, + 0.003733946243301034, + 0.0033850206527858973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.01109242532402277 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643004111945629 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643004111945629, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643004111945629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13316330313682556, + 0.046261612325906754, + 0.015323345549404621, + 0.09597186744213104, + 0.3379864990711212, + 0.2904880940914154, + 0.030034605413675308, + 0.013063671998679638, + 0.02129656821489334, + 0.01641043648123741 + ], + "credit_map": [ + 0.0073727769777178764, + 0.007268539629876614, + 0.02938828058540821, + 0.01894466206431389, + 0.009452132508158684, + 0.021111709997057915, + 0.0112885981798172, + 0.02833660878241062, + 0.00427889171987772, + 0.007120042573660612, + 0.005424624774605036, + 0.007552471477538347, + 0.00660166097804904, + 0.004142066463828087, + 0.0062696123495697975, + 0.004872245714068413, + 0.0018186629749834538, + 0.0016786458436399698, + 0.001779023208655417, + 0.0017149008344858885, + 0.0014354559825733304, + 0.002353158313781023, + 0.002724163932725787, + 0.001819335506297648, + 0.010370105504989624, + 0.017564695328474045, + 0.007667188066989183, + 0.016547219827771187, + 0.011270117945969105, + 0.008881825022399426, + 0.015811409801244736, + 0.007859311066567898, + 0.01683991588652134, + 0.037340205162763596, + 0.00951799750328064, + 0.01689019426703453, + 0.11623768508434296, + 0.10126741230487823, + 0.022322578355669975, + 0.01757049933075905, + 0.01852557063102722, + 0.04030627757310867, + 0.01980126090347767, + 0.0753476694226265, + 0.05265776440501213, + 0.023738734424114227, + 0.04076019674539566, + 0.01935061812400818, + 0.0023030589800328016, + 0.004943872801959515, + 0.0023290892131626606, + 0.0038411689456552267, + 0.005194618366658688, + 0.0036411823239177465, + 0.0048447418957948685, + 0.0029368740506470203, + 0.0013198343804106116, + 0.00085838558152318, + 0.001971010584384203, + 0.001453031087294221, + 0.0013332272646948695, + 0.0015999901806935668, + 0.0026816572062671185, + 0.0018465372268110514, + 0.0037202818784862757, + 0.0025408733636140823, + 0.0028084551449865103, + 0.0022593478206545115, + 0.0033627874217927456, + 0.0015223841182887554, + 0.003039472969248891, + 0.002042967127636075, + 0.00205327314324677, + 0.0018191840499639511, + 0.0025979673955589533, + 0.0018697070190683007, + 0.0020581320859491825, + 0.001376852858811617, + 0.0025328448973596096, + 0.0021024758461862803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.008643004111945629 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989026010036469 + ], + "opposite_logprob_tokens": [ + -1.8048902750015259 + ], + "logprob_reward": -0.17989026010036469, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989026010036469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10464354604482651, + 0.04919935762882233, + 0.02636767365038395, + 0.12281762063503265, + 0.2948942482471466, + 0.2698827087879181, + 0.02514358051121235, + 0.021336419507861137, + 0.05322285741567612, + 0.03249197453260422 + ], + "credit_map": [ + 0.008022520691156387, + 0.015155418775975704, + 0.014049828052520752, + 0.0152797382324934, + 0.009629379957914352, + 0.014283902943134308, + 0.012619756162166595, + 0.015603006817400455, + 0.005284623242914677, + 0.006602243520319462, + 0.007194026373326778, + 0.00566752627491951, + 0.006647361908107996, + 0.006110539194196463, + 0.005621273070573807, + 0.006071766838431358, + 0.002471351530402899, + 0.0037909371312707663, + 0.0036384896375238895, + 0.003574077971279621, + 0.003219708101823926, + 0.003862044308334589, + 0.003414296777918935, + 0.002396768657490611, + 0.007605415303260088, + 0.0082851005718112, + 0.009310738183557987, + 0.01575535163283348, + 0.012612525373697281, + 0.007933415472507477, + 0.034645676612854004, + 0.026669394224882126, + 0.015127381309866905, + 0.03988305851817131, + 0.032780371606349945, + 0.03687829524278641, + 0.05838072672486305, + 0.03679642826318741, + 0.03394481912255287, + 0.04110318049788475, + 0.02112867496907711, + 0.01383985299617052, + 0.018966129049658775, + 0.10737617313861847, + 0.023018859326839447, + 0.011830640956759453, + 0.05642685294151306, + 0.017295511439442635, + 0.0025724435690790415, + 0.003596722846850753, + 0.0022567834239453077, + 0.0029393937438726425, + 0.003714685793966055, + 0.0026717153377830982, + 0.003450509160757065, + 0.003941325005143881, + 0.0021397771779447794, + 0.0017922287806868553, + 0.0034176530316472054, + 0.003036698093637824, + 0.0028151816222816706, + 0.001387918135151267, + 0.0030322198290377855, + 0.003714742138981819, + 0.011317746713757515, + 0.005927837919443846, + 0.009493404068052769, + 0.008045931346714497, + 0.003951089922338724, + 0.003614535555243492, + 0.005367327015846968, + 0.005504986736923456, + 0.005531100556254387, + 0.003485811874270439, + 0.003908218815922737, + 0.005487706977874041, + 0.003837556578218937, + 0.002394456649199128, + 0.0036100067663937807, + 0.004237114917486906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.17989026010036469 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091939423233271 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091939423233271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091939423233271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12184689193964005, + 0.05803379788994789, + 0.02217165008187294, + 0.30412736535072327, + 0.1552451252937317, + 0.20339351892471313, + 0.016323912888765335, + 0.022321084514260292, + 0.05948767438530922, + 0.037048954516649246 + ], + "credit_map": [ + 0.006666913628578186, + 0.006841967348009348, + 0.026138542219996452, + 0.018514135852456093, + 0.00838051550090313, + 0.01760294660925865, + 0.011120992712676525, + 0.026580870151519775, + 0.005858106072992086, + 0.0037661355454474688, + 0.006206257734447718, + 0.0063104042783379555, + 0.00990715529769659, + 0.010180062614381313, + 0.005326590966433287, + 0.010479085147380829, + 0.002656378783285618, + 0.002858043648302555, + 0.0023870915174484253, + 0.0025772019289433956, + 0.0025199425872415304, + 0.004502457100898027, + 0.0027442483697086573, + 0.0019262864952906966, + 0.018682606518268585, + 0.014913675375282764, + 0.025068074464797974, + 0.031162528321146965, + 0.024908864870667458, + 0.020437654107809067, + 0.11102016270160675, + 0.05793380364775658, + 0.007701114751398563, + 0.021068179979920387, + 0.00495509710162878, + 0.01240313146263361, + 0.05088980868458748, + 0.03964652493596077, + 0.010616017505526543, + 0.00796525739133358, + 0.014358273707330227, + 0.025205817073583603, + 0.018451154232025146, + 0.05941559746861458, + 0.022677449509501457, + 0.011586813256144524, + 0.031702347099781036, + 0.019996073096990585, + 0.0015274479519575834, + 0.002217530505731702, + 0.001514355419203639, + 0.0016119269421324134, + 0.002625212538987398, + 0.002562959911301732, + 0.002837544772773981, + 0.0014269347302615643, + 0.001959032379090786, + 0.001501782680861652, + 0.00331529276445508, + 0.002010378986597061, + 0.003672410733997822, + 0.0027198863681405783, + 0.0028275817167013884, + 0.004314718302339315, + 0.013276830315589905, + 0.005573958624154329, + 0.011528287082910538, + 0.009804407134652138, + 0.0034937963355332613, + 0.00324888713657856, + 0.006569170858711004, + 0.0059923334047198296, + 0.007101158611476421, + 0.003724323585629463, + 0.005145343020558357, + 0.00521075027063489, + 0.0029778918251395226, + 0.0025255209766328335, + 0.005237285513430834, + 0.0051266783848404884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.004091939423233271 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a panoramic shot rendered in a polished Pixar style, the camera glides smoothly in an arc around a central, highly polished, dark wooden stage. At its center, the joyful taiko character, Wadadou, performs a playful dance. His body, a perfectly spherical light brown drum with subtle textures, gleams under the bright, warm stage lights. A dazzling cartoon smile perpetually etched on his face, flanked by two round, rosy cheeks, while his large, simple black eyes sparkle with pure joy. He bounces rhythmically on his base, his entire round body swaying slightly with each jump, while his short, slender, cylindrical limbs pulsate with exuberant, carefree energy. The spotlessly polished stage floor reflects his every movement in a soft, vibrant color. Against the vibrant, soft-focus background, colorful hanging banners and glowing paper lanterns fill the festive world, creating a scene awash in saturated, warm colors. The entire image is bathed in a soft, celebratory light that eliminates harsh shadows, creating an overwhelmingly joyous and energetic atmosphere of pure, infectious happiness.", + "seed": 28286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28286_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": -0.022726147435605526, + "vlm_logprob": -0.022726147435605526, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.022726147435605526, + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06276921182870865 + ], + "opposite_logprob_tokens": [ + -3.0627691745758057 + ], + "logprob_reward": -0.06276921182870865, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06276921182870865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1958829015493393, + 0.049247048795223236, + 0.07836209237575531, + 0.1687074452638626, + 0.04862137883901596, + 0.05926039442420006, + 0.163215771317482, + 0.07375621050596237, + 0.08590640127658844, + 0.07704033702611923 + ], + "credit_map": [ + 0.014312054961919785, + 0.016171161085367203, + 0.024254655465483665, + 0.04850132018327713, + 0.0145800169557333, + 0.020048795267939568, + 0.03601468726992607, + 0.022000210359692574, + 0.0043268585577607155, + 0.0038313816767185926, + 0.005212017334997654, + 0.0085869450122118, + 0.004380660131573677, + 0.004947794135659933, + 0.011990172788500786, + 0.005971218924969435, + 0.006996558979153633, + 0.009477741084992886, + 0.0034465352073311806, + 0.008026177063584328, + 0.019324805587530136, + 0.019659243524074554, + 0.0065050553530454636, + 0.004925974644720554, + 0.017230555415153503, + 0.013658967800438404, + 0.00929528009146452, + 0.01031040120869875, + 0.035359982401132584, + 0.0492599792778492, + 0.01074314583092928, + 0.022849133238196373, + 0.005151531659066677, + 0.011483212001621723, + 0.005616762209683657, + 0.0066623380407691, + 0.005847719497978687, + 0.004439983982592821, + 0.005735259503126144, + 0.0036845672875642776, + 0.006475331727415323, + 0.005183208733797073, + 0.011849943548440933, + 0.009413962252438068, + 0.002808086108416319, + 0.01031043753027916, + 0.005947684869170189, + 0.0072717368602752686, + 0.016835469752550125, + 0.029179949313402176, + 0.011614933609962463, + 0.0358247309923172, + 0.01666269451379776, + 0.02022242359817028, + 0.02094327285885811, + 0.011932301335036755, + 0.006494990549981594, + 0.010094805620610714, + 0.01051203440874815, + 0.008596043102443218, + 0.008829007856547832, + 0.011079158633947372, + 0.00945069920271635, + 0.00869947113096714, + 0.005164962261915207, + 0.010987001471221447, + 0.010945546440780163, + 0.023202555254101753, + 0.009974348358809948, + 0.005789852235466242, + 0.012788528576493263, + 0.007053602486848831, + 0.00906779058277607, + 0.0071240197867155075, + 0.011795314960181713, + 0.00996831152588129, + 0.00920830573886633, + 0.007729372475296259, + 0.01398996077477932, + 0.00815726350992918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.06276921182870865 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265297368168831 + ], + "opposite_logprob_tokens": [ + -5.255265235900879 + ], + "logprob_reward": -0.005265297368168831, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265297368168831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3126085698604584, + 0.06359627842903137, + 0.14583440124988556, + 0.08201885223388672, + 0.03674362227320671, + 0.0746147409081459, + 0.08615484088659286, + 0.11079003661870956, + 0.025158531963825226, + 0.062480151653289795 + ], + "credit_map": [ + 0.01744701899588108, + 0.02086082473397255, + 0.04314141348004341, + 0.089832603931427, + 0.015085970051586628, + 0.025407861918210983, + 0.05136715620756149, + 0.04946571961045265, + 0.007610656321048737, + 0.008496727794408798, + 0.006134177558124065, + 0.006919109728187323, + 0.011759160086512566, + 0.0072745284996926785, + 0.0067916628904640675, + 0.008610261604189873, + 0.010030818171799183, + 0.012346025556325912, + 0.005116940010339022, + 0.006956177297979593, + 0.042368970811367035, + 0.053116172552108765, + 0.008741588331758976, + 0.0071576968766748905, + 0.006928873248398304, + 0.009175173006951809, + 0.002798582660034299, + 0.005316395778208971, + 0.026423979550600052, + 0.022175047546625137, + 0.004568541422486305, + 0.004632262513041496, + 0.0033152943942695856, + 0.006351512856781483, + 0.003342488780617714, + 0.0053994981572031975, + 0.004383409861475229, + 0.006929760333150625, + 0.003818920347839594, + 0.0032027363777160645, + 0.008904466405510902, + 0.0074943965300917625, + 0.01338729914277792, + 0.009924403391778469, + 0.006319637410342693, + 0.011970586143434048, + 0.007357277907431126, + 0.009256665594875813, + 0.012845694087445736, + 0.01789577677845955, + 0.005848419386893511, + 0.010540660470724106, + 0.009280718863010406, + 0.014824705198407173, + 0.007013555150479078, + 0.007905310951173306, + 0.009793766774237156, + 0.020882315933704376, + 0.01310482807457447, + 0.020502524450421333, + 0.012180674821138382, + 0.01110673788934946, + 0.01203941646963358, + 0.011179781518876553, + 0.0020074755884706974, + 0.002882398897781968, + 0.0032874594908207655, + 0.0047818454913794994, + 0.003373943967744708, + 0.0031277998350560665, + 0.003646950935944915, + 0.002050658455118537, + 0.006157975643873215, + 0.008658202365040779, + 0.008303279988467693, + 0.006365188397467136, + 0.009038684889674187, + 0.011478351429104805, + 0.0074044642969965935, + 0.005074003711342812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.005265297368168831 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537001632153988 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537001632153988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537001632153988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19942566752433777, + 0.046849943697452545, + 0.18034785985946655, + 0.0633968710899353, + 0.05715559422969818, + 0.08023468405008316, + 0.08551962673664093, + 0.10979726910591125, + 0.1378234326839447, + 0.039449069648981094 + ], + "credit_map": [ + 0.012812141329050064, + 0.021795598790049553, + 0.025333192199468613, + 0.0461878627538681, + 0.018366266041994095, + 0.020551828667521477, + 0.026465125381946564, + 0.027913644909858704, + 0.0049897111020982265, + 0.00555158406496048, + 0.0041665672324597836, + 0.004998830147087574, + 0.00920171570032835, + 0.007352988235652447, + 0.00485779857262969, + 0.005730750970542431, + 0.017497582361102104, + 0.008855562657117844, + 0.004762735217809677, + 0.00756016606464982, + 0.04421543702483177, + 0.08108081668615341, + 0.006522173993289471, + 0.009853394702076912, + 0.006942093372344971, + 0.006337939761579037, + 0.004110571462661028, + 0.008531449362635612, + 0.005106039810925722, + 0.01316159963607788, + 0.008013470098376274, + 0.011193706654012203, + 0.004126125015318394, + 0.00977801252156496, + 0.00595507025718689, + 0.01235870085656643, + 0.006616791244596243, + 0.00380867812782526, + 0.009918437339365482, + 0.004593780264258385, + 0.007850619964301586, + 0.005445247516036034, + 0.02106579951941967, + 0.00824418943375349, + 0.003532869042828679, + 0.024495882913470268, + 0.004963760729879141, + 0.004636312369257212, + 0.011068976484239101, + 0.013019477017223835, + 0.008279326371848583, + 0.013078917749226093, + 0.009185031056404114, + 0.012148476205766201, + 0.008409341797232628, + 0.010330076329410076, + 0.011180098168551922, + 0.016872558742761612, + 0.01563810557126999, + 0.01576649770140648, + 0.016761578619480133, + 0.010303293354809284, + 0.013055778108537197, + 0.010219356045126915, + 0.007548669818788767, + 0.017474787309765816, + 0.018902521580457687, + 0.03906790167093277, + 0.014150123111903667, + 0.009104609489440918, + 0.01933717541396618, + 0.012237649410963058, + 0.003934602718800306, + 0.005498688202351332, + 0.004636628553271294, + 0.004406305495649576, + 0.005906306207180023, + 0.005445913411676884, + 0.006074816919863224, + 0.003545811166986823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.012537001632153988 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287026211619377 + ], + "opposite_logprob_tokens": [ + -3.773287057876587 + ], + "logprob_reward": -0.023287026211619377, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287026211619377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10021212697029114, + 0.04419670253992081, + 0.19513024389743805, + 0.08743597567081451, + 0.034805264323949814, + 0.09800326824188232, + 0.09518072009086609, + 0.08415627479553223, + 0.16510851681232452, + 0.09577088057994843 + ], + "credit_map": [ + 0.006429826840758324, + 0.008599186316132545, + 0.011044783517718315, + 0.026096398010849953, + 0.00794309750199318, + 0.009798189625144005, + 0.01637149043381214, + 0.013929151929914951, + 0.005269268993288279, + 0.0053812055848538876, + 0.0029667646158486605, + 0.0056651895865798, + 0.007320389151573181, + 0.00476386584341526, + 0.006607874296605587, + 0.006222141906619072, + 0.01624966412782669, + 0.009233053773641586, + 0.00561847910284996, + 0.006020409986376762, + 0.052456725388765335, + 0.08874277025461197, + 0.007217947859317064, + 0.009591194801032543, + 0.007155261468142271, + 0.00543432030826807, + 0.002620407147333026, + 0.0032395857851952314, + 0.02177804335951805, + 0.03074386529624462, + 0.004504249896854162, + 0.01196025125682354, + 0.003059288254007697, + 0.006417674478143454, + 0.00544198602437973, + 0.005869855172932148, + 0.003951970022171736, + 0.0032919165678322315, + 0.003988344222307205, + 0.0027842288836836815, + 0.010301297530531883, + 0.005903720855712891, + 0.022981345653533936, + 0.007026333827525377, + 0.004006251227110624, + 0.034622371196746826, + 0.005948927253484726, + 0.007213023956865072, + 0.011781381443142891, + 0.013865644112229347, + 0.008969361893832684, + 0.025570567697286606, + 0.007254455704241991, + 0.01022154837846756, + 0.009123768657445908, + 0.008393991738557816, + 0.006826252676546574, + 0.00978299230337143, + 0.010638792999088764, + 0.009483558125793934, + 0.008961444720625877, + 0.010737085714936256, + 0.015004691667854786, + 0.01272144727408886, + 0.010872182436287403, + 0.011235849931836128, + 0.023567691445350647, + 0.04602094367146492, + 0.014531021937727928, + 0.010817714966833591, + 0.02881690114736557, + 0.01924620196223259, + 0.011249682866036892, + 0.011362091638147831, + 0.015455939806997776, + 0.011628435924649239, + 0.014686414040625095, + 0.010672951117157936, + 0.011695523746311665, + 0.009019844233989716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.023287026211619377 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772200137376785 + ], + "opposite_logprob_tokens": [ + -4.634772300720215 + ], + "logprob_reward": -0.009772200137376785, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772200137376785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17440608143806458, + 0.04932620748877525, + 0.19007061421871185, + 0.05054846778512001, + 0.026119772344827652, + 0.2522989511489868, + 0.03683149069547653, + 0.13657985627651215, + 0.018909236416220665, + 0.06490933150053024 + ], + "credit_map": [ + 0.008850498124957085, + 0.009784800000488758, + 0.02771340310573578, + 0.04628939554095268, + 0.011086280457675457, + 0.014109847135841846, + 0.03169156610965729, + 0.02488028258085251, + 0.005292433779686689, + 0.005080219823867083, + 0.005280487239360809, + 0.00710415979847312, + 0.008497742004692554, + 0.006062301807105541, + 0.005762265995144844, + 0.006246602162718773, + 0.012458295561373234, + 0.011835819110274315, + 0.0077331638894975185, + 0.012623809278011322, + 0.04980195686221123, + 0.07240971177816391, + 0.013448808342218399, + 0.009759029373526573, + 0.004369438625872135, + 0.004677793476730585, + 0.002822879236191511, + 0.007439173758029938, + 0.011248892173171043, + 0.0063669742085039616, + 0.009178761392831802, + 0.004444556310772896, + 0.002785859163850546, + 0.004599318839609623, + 0.0040959385223686695, + 0.003577193710952997, + 0.0029379818588495255, + 0.0030860297847539186, + 0.0034012312535196543, + 0.0016362187452614307, + 0.021115994080901146, + 0.014494013972580433, + 0.07497404515743256, + 0.017887206748127937, + 0.009290286339819431, + 0.08545778691768646, + 0.014043827541172504, + 0.015035796910524368, + 0.0037774350494146347, + 0.007442471571266651, + 0.003140043467283249, + 0.006084332708269358, + 0.0049787419848144054, + 0.004629821516573429, + 0.004283033777028322, + 0.0024956108536571264, + 0.014831592328846455, + 0.01582244038581848, + 0.02336695045232773, + 0.014078844338655472, + 0.013465926051139832, + 0.016278116032481194, + 0.018103519454598427, + 0.02063247188925743, + 0.0017171159852296114, + 0.0030706552788615227, + 0.001819336786866188, + 0.003241023514419794, + 0.0020143617875874043, + 0.0022524804808199406, + 0.002820131601765752, + 0.001974131679162383, + 0.008086171001195908, + 0.006775142624974251, + 0.011086110025644302, + 0.005758165381848812, + 0.005869022570550442, + 0.009160582907497883, + 0.009433492086827755, + 0.008740643970668316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.009772200137376785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a panoramic shot rendered in a polished Pixar style, the camera glides smoothly in an arc around a central, highly polished, dark wooden stage. At its center, the joyful taiko character, Wadadou, performs a playful dance. His body, a perfectly spherical light brown drum with subtle textures, gleams under the bright, warm stage lights. A dazzling cartoon smile perpetually etched on his face, flanked by two round, rosy cheeks, while his large, simple black eyes sparkle with pure joy. He bounces rhythmically on his base, his entire round body swaying slightly with each jump, while his short, slender, cylindrical limbs pulsate with exuberant, carefree energy. The spotlessly polished stage floor reflects his every movement in a soft, vibrant color. Against the vibrant, soft-focus background, colorful hanging banners and glowing paper lanterns fill the festive world, creating a scene awash in saturated, warm colors. The entire image is bathed in a soft, celebratory light that eliminates harsh shadows, creating an overwhelmingly joyous and energetic atmosphere of pure, infectious happiness.", + "seed": 28287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_13/index_9776_seed_28287_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": -0.019311734614893793, + "vlm_logprob": -0.019311734614893793, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.019311734614893793, + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05414087697863579 + ], + "opposite_logprob_tokens": [ + -3.17914080619812 + ], + "logprob_reward": -0.05414087697863579, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05414087697863579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3870846927165985, + 0.13342225551605225, + 0.054975032806396484, + 0.02874261699616909, + 0.07139883190393448, + 0.06323302537202835, + 0.044555217027664185, + 0.06335800141096115, + 0.08879793435335159, + 0.06443235278129578 + ], + "credit_map": [ + 0.040103472769260406, + 0.016710616648197174, + 0.015697533264756203, + 0.027404513210058212, + 0.01924741081893444, + 0.02572634629905224, + 0.11062777787446976, + 0.13156703114509583, + 0.009241782128810883, + 0.013109934516251087, + 0.016884559765458107, + 0.015113978646695614, + 0.015599686652421951, + 0.018230436369776726, + 0.025063427165150642, + 0.02017844468355179, + 0.006362477317452431, + 0.0041913362219929695, + 0.008946836926043034, + 0.008028155192732811, + 0.004606963135302067, + 0.009123158641159534, + 0.005910609383136034, + 0.00780549505725503, + 0.0030914631206542253, + 0.002865066286176443, + 0.004091926850378513, + 0.0032877232879400253, + 0.0032787469681352377, + 0.003066117176786065, + 0.0038718588184565306, + 0.005189714953303337, + 0.006034730467945337, + 0.009190520271658897, + 0.01267640758305788, + 0.01750330626964569, + 0.007593631278723478, + 0.0038231874350458384, + 0.007184957154095173, + 0.007392100524157286, + 0.008052730932831764, + 0.006612157449126244, + 0.01035112515091896, + 0.013033038936555386, + 0.004493162035942078, + 0.005635485984385014, + 0.009491488337516785, + 0.005563842132687569, + 0.008058728650212288, + 0.004977110307663679, + 0.007733146660029888, + 0.004879825282841921, + 0.0053822556510567665, + 0.004197552800178528, + 0.004907730035483837, + 0.004418864380568266, + 0.009897373616695404, + 0.006663862615823746, + 0.009798695333302021, + 0.00624814024195075, + 0.003850929206237197, + 0.00760524021461606, + 0.008012802340090275, + 0.011280958540737629, + 0.01054807286709547, + 0.010245286859571934, + 0.014516796916723251, + 0.015866778790950775, + 0.0069073704071342945, + 0.005412580445408821, + 0.012916543520987034, + 0.012384508736431599, + 0.004183144774287939, + 0.00610628305003047, + 0.008743948303163052, + 0.012401009909808636, + 0.005173468496650457, + 0.0085262730717659, + 0.009533623233437538, + 0.009764604270458221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.05414087697863579 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648474045097828 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648474045097828, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648474045097828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35251206159591675, + 0.1251472532749176, + 0.051119767129421234, + 0.031881183385849, + 0.09828915446996689, + 0.04422317072749138, + 0.09413426369428635, + 0.10352775454521179, + 0.06286370009183884, + 0.03630170598626137 + ], + "credit_map": [ + 0.038457050919532776, + 0.021632753312587738, + 0.02093612775206566, + 0.023110924288630486, + 0.03300910443067551, + 0.04021647945046425, + 0.07789375633001328, + 0.0972558781504631, + 0.00884136650711298, + 0.011188068427145481, + 0.011368000879883766, + 0.01388594787567854, + 0.01452281791716814, + 0.019675586372613907, + 0.02603239193558693, + 0.01963307522237301, + 0.005639289040118456, + 0.0033663141075521708, + 0.006160943303257227, + 0.007846161723136902, + 0.006770758423954248, + 0.007872963324189186, + 0.0056604985147714615, + 0.007802838459610939, + 0.0033350714948028326, + 0.002926028333604336, + 0.003668322926387191, + 0.004318796098232269, + 0.00540096266195178, + 0.0028715687803924084, + 0.004441073164343834, + 0.004919358994811773, + 0.006138141732662916, + 0.007162187248468399, + 0.019651608541607857, + 0.024005750194191933, + 0.006251819897443056, + 0.005650069564580917, + 0.01867108978331089, + 0.010758493095636368, + 0.003842298872768879, + 0.003844192251563072, + 0.008368163369596004, + 0.008428915403783321, + 0.0037830618675798178, + 0.0031702963169664145, + 0.008939391002058983, + 0.0038468504790216684, + 0.02034968137741089, + 0.007018068805336952, + 0.014573213644325733, + 0.011314036324620247, + 0.006527982652187347, + 0.014942740090191364, + 0.010439322330057621, + 0.008969219401478767, + 0.012521017342805862, + 0.012656785547733307, + 0.014058028347790241, + 0.012673069722950459, + 0.010588400065898895, + 0.011805192567408085, + 0.01187423150986433, + 0.017351027578115463, + 0.004905974958091974, + 0.00920333806425333, + 0.008045025169849396, + 0.011511704884469509, + 0.0066741169430315495, + 0.0038296710699796677, + 0.010158572345972061, + 0.008535302244126797, + 0.0026295834686607122, + 0.004746951628476381, + 0.003981590736657381, + 0.0058456724509596825, + 0.003796014469116926, + 0.0056114960461854935, + 0.005002654157578945, + 0.004687742330133915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004648474045097828 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43118080496788025, + 0.0828380212187767, + 0.03745746240019798, + 0.025117583572864532, + 0.031320374459028244, + 0.04753495752811432, + 0.09814237058162689, + 0.15648768842220306, + 0.06321144104003906, + 0.026709282770752907 + ], + "credit_map": [ + 0.034883998334407806, + 0.022054824978113174, + 0.015430349856615067, + 0.02814961038529873, + 0.03342777490615845, + 0.04562543332576752, + 0.11521721631288528, + 0.13639162480831146, + 0.008307453244924545, + 0.007068471983075142, + 0.014152324758470058, + 0.007674625143408775, + 0.00731425266712904, + 0.01274601835757494, + 0.013408638536930084, + 0.012166241183876991, + 0.00587820028886199, + 0.0028372076340019703, + 0.005679825320839882, + 0.005817818455398083, + 0.0026082315016537905, + 0.005185231566429138, + 0.004485627636313438, + 0.004965323023498058, + 0.00265310937538743, + 0.001994529040530324, + 0.003554415190592408, + 0.003925851546227932, + 0.004310694523155689, + 0.0016742530278861523, + 0.0036825796123594046, + 0.003322153352200985, + 0.002165780868381262, + 0.0032120102550834417, + 0.005917742382735014, + 0.007194759789854288, + 0.002241553505882621, + 0.0019061820348724723, + 0.0055273897014558315, + 0.0031549567356705666, + 0.0031734106596559286, + 0.003435047809034586, + 0.009910725057125092, + 0.012695427983999252, + 0.002115722047165036, + 0.0020262303296476603, + 0.011297951452434063, + 0.002880444983020425, + 0.01860838010907173, + 0.008505906909704208, + 0.014857886359095573, + 0.010891314595937729, + 0.007532664109021425, + 0.011454736813902855, + 0.01651168242096901, + 0.009779801592230797, + 0.024430060759186745, + 0.015743933618068695, + 0.023151662200689316, + 0.01776854321360588, + 0.008345995098352432, + 0.01650240458548069, + 0.028210971504449844, + 0.02233411930501461, + 0.005544643383473158, + 0.007971256971359253, + 0.007809772621840239, + 0.012071981094777584, + 0.007199838757514954, + 0.00479700556024909, + 0.00976711418479681, + 0.008049828931689262, + 0.0020858703646808863, + 0.003530836198478937, + 0.002564712893217802, + 0.003963778261095285, + 0.003915834240615368, + 0.0029572374187409878, + 0.00393515033647418, + 0.0037558660842478275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.012542181648314 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288888856768608 + ], + "opposite_logprob_tokens": [ + -3.7732889652252197 + ], + "logprob_reward": -0.023288888856768608, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288888856768608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40920937061309814, + 0.04255734384059906, + 0.01577736623585224, + 0.023259295150637627, + 0.013793347403407097, + 0.026694882661104202, + 0.06444921344518661, + 0.1661357432603836, + 0.07719486951828003, + 0.16092857718467712 + ], + "credit_map": [ + 0.03983182832598686, + 0.012346597388386726, + 0.012062097899615765, + 0.012639817781746387, + 0.01842203550040722, + 0.024338873103260994, + 0.1323307603597641, + 0.15723735094070435, + 0.003796167206019163, + 0.003997007850557566, + 0.005228192079812288, + 0.004446416161954403, + 0.005738411098718643, + 0.007307381834834814, + 0.006244864780455828, + 0.005798902828246355, + 0.0020207406487315893, + 0.000943697348702699, + 0.0017358377808704972, + 0.0032345892395824194, + 0.0011309469118714333, + 0.0020293178968131542, + 0.002586649265140295, + 0.002095585921779275, + 0.0022743360605090857, + 0.0018513737013563514, + 0.0037733926437795162, + 0.0030377162620425224, + 0.004607142880558968, + 0.0017334966687485576, + 0.0030288349371403456, + 0.002953000133857131, + 0.001088786986656487, + 0.0017697723815217614, + 0.0022230518516153097, + 0.0024275670293718576, + 0.0010348388459533453, + 0.001081343973055482, + 0.002050104085355997, + 0.002117882715538144, + 0.002603275701403618, + 0.002814439358189702, + 0.004876624792814255, + 0.005998932756483555, + 0.0020011835731565952, + 0.0012914901599287987, + 0.005396705120801926, + 0.0017122322460636497, + 0.010775351896882057, + 0.0053187161684036255, + 0.010284172371029854, + 0.008122815750539303, + 0.004945423454046249, + 0.007586278021335602, + 0.010904445312917233, + 0.006512012332677841, + 0.019819142296910286, + 0.010895545594394207, + 0.022516610100865364, + 0.02154579386115074, + 0.008936751633882523, + 0.01771114394068718, + 0.03612697869539261, + 0.02858377806842327, + 0.0075839972123503685, + 0.008774268440902233, + 0.010926981456577778, + 0.014700109139084816, + 0.0059075490571558475, + 0.0048840283416211605, + 0.013491137884557247, + 0.010926799848675728, + 0.009135348722338676, + 0.026774859055876732, + 0.013478518463671207, + 0.028409119695425034, + 0.017062287777662277, + 0.01436619833111763, + 0.031604915857315063, + 0.020097333937883377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.023288888856768608 + }, + { + "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 panoramic and gliding smoothly in an arc around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a spherical, light brown taiko drum character named Wadadou?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character performing a playful dance by bouncing rhythmically on its base with swaying motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a highly polished dark wooden stage with a floor that reflects the character's movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background filled with colorful hanging banners and glowing paper lanterns in a soft-focus style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019382515456527472 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019382515456527472, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019382515456527472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3258727788925171, + 0.0881127193570137, + 0.04192773997783661, + 0.05266847461462021, + 0.0413113608956337, + 0.041494593024253845, + 0.19666051864624023, + 0.10963460057973862, + 0.029062960296869278, + 0.0732542872428894 + ], + "credit_map": [ + 0.03330772742629051, + 0.011531870812177658, + 0.014284004457294941, + 0.01252909004688263, + 0.014854851178824902, + 0.021448027342557907, + 0.09811171889305115, + 0.1198054850101471, + 0.007398535497486591, + 0.007443714886903763, + 0.013873140327632427, + 0.008949967101216316, + 0.012129422277212143, + 0.010071951895952225, + 0.01641608215868473, + 0.011829904280602932, + 0.005688031204044819, + 0.004374416079372168, + 0.005251866299659014, + 0.0053685978055000305, + 0.0040437146089971066, + 0.00697297090664506, + 0.004485812038183212, + 0.005742328241467476, + 0.004621448926627636, + 0.0045663705095648766, + 0.009920204058289528, + 0.005138487555086613, + 0.008577636443078518, + 0.00470141414552927, + 0.0059842634946107864, + 0.009158650413155556, + 0.003385961288586259, + 0.005086544901132584, + 0.006975402124226093, + 0.008698254823684692, + 0.003646636614575982, + 0.002924165455624461, + 0.00527085131034255, + 0.005323542281985283, + 0.0038132050540298223, + 0.004221160896122456, + 0.006753881927579641, + 0.00630649458616972, + 0.0032334006391465664, + 0.004907272290438414, + 0.006342380773276091, + 0.005916797090321779, + 0.04954664781689644, + 0.012449176050722599, + 0.03351591154932976, + 0.021899662911891937, + 0.012751737609505653, + 0.02858569845557213, + 0.02102987840771675, + 0.016881797462701797, + 0.01582726649940014, + 0.007341980002820492, + 0.017428124323487282, + 0.015084020793437958, + 0.0051603675819933414, + 0.013837396167218685, + 0.01931813731789589, + 0.01563730463385582, + 0.0036723988596349955, + 0.003254694165661931, + 0.005097200162708759, + 0.0036568089853972197, + 0.0028326562605798244, + 0.0026295066345483065, + 0.003340956522151828, + 0.00457873847335577, + 0.006833680439740419, + 0.009651033207774162, + 0.011669705621898174, + 0.012067875824868679, + 0.005646619014441967, + 0.008732878603041172, + 0.006665002088993788, + 0.011987493373453617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0019382515456527472 + } + ], + "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/0000060/group_14/index_tensor([6621, 6621])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_tensor([6621, 6621])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a69e01a5b276bc9de6d88083d92ed9bef6864270 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_tensor([6621, 6621])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly low-angle close-up, shot with a sleek, doll-like backward tracking angle, perfectly captures the determined face of an East Asian woman in her twenties as she steadily moves forward. Her brow is slightly furrowed in concentration, her dark, focused eyes fixed on a single point off-screen, her lips pressed tightly together. Her sleek, shoulder-length black hair, slightly damp from the cool night air, captures the ever-changing ambient light. She wears a charcoal gray turtleneck windbreaker, its waterproof fabric reflecting a kaleidoscopic hue. The nighttime cityscape behind her becomes a vibrant, abstract tapestry composed of motion-blurred streaks—electric blue, neon pink, and warm yellow from traffic and storefronts—with fleeting highlights traced across her sharp cheekbones and jawline. Her purposeful walk is relentless and steady, her head held level, creating a powerful sense of unwavering determination and forward momentum amid the chaotic, out-of-focus beauty of the urban environment, establishing an atmosphere of intense loneliness.", + "seed": 27851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_6621_seed_27851_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": -0.0380405968055129, + "vlm_logprob": -0.0380405968055129, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.0380405968055129, + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009600185789167881 + ], + "opposite_logprob_tokens": [ + -4.8846001625061035 + ], + "logprob_reward": -0.009600185789167881, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009600185789167881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16373594105243683, + 0.33321788907051086, + 0.03733142837882042, + 0.03814917430281639, + 0.1194186732172966, + 0.06713970005512238, + 0.028362741693854332, + 0.13628172874450684, + 0.01940179616212845, + 0.05696093663573265 + ], + "credit_map": [ + 0.01945812813937664, + 0.016450608149170876, + 0.012647422961890697, + 0.02445572428405285, + 0.018383730202913284, + 0.03051542490720749, + 0.01580885984003544, + 0.026016050949692726, + 0.04917870834469795, + 0.04299022629857063, + 0.03093588538467884, + 0.1032223105430603, + 0.018809303641319275, + 0.035686101764440536, + 0.028897518292069435, + 0.023497823625802994, + 0.0034981449134647846, + 0.0048125130124390125, + 0.0047471183352172375, + 0.004402025602757931, + 0.0034730785991996527, + 0.004002989735454321, + 0.008061143569648266, + 0.004334417171776295, + 0.0042114295065402985, + 0.003961094189435244, + 0.004379586316645145, + 0.0039484235458076, + 0.006005391478538513, + 0.004577824380248785, + 0.005365851800888777, + 0.005699573550373316, + 0.007465590722858906, + 0.009555854834616184, + 0.02039591781795025, + 0.014860287308692932, + 0.011055962182581425, + 0.019684433937072754, + 0.013373149558901787, + 0.023027483373880386, + 0.003263230202719569, + 0.006538597401231527, + 0.005321054719388485, + 0.00852117221802473, + 0.011138717643916607, + 0.010487081483006477, + 0.013891076669096947, + 0.007978766225278378, + 0.0026177598629146814, + 0.003023662604391575, + 0.004867421463131905, + 0.003452796023339033, + 0.003383859759196639, + 0.004099163226783276, + 0.003180644242092967, + 0.0037374349776655436, + 0.006457681767642498, + 0.010955704376101494, + 0.011938882060348988, + 0.014057828113436699, + 0.01300546433776617, + 0.03182154893875122, + 0.021441666409373283, + 0.026602940633893013, + 0.0023797855246812105, + 0.0020824011880904436, + 0.002113131107762456, + 0.0019301308784633875, + 0.0032453115563839674, + 0.002886393340304494, + 0.0024361517280340195, + 0.0023284906055778265, + 0.006640973500907421, + 0.0057516577653586864, + 0.010082336142659187, + 0.008428932167589664, + 0.0073048630729317665, + 0.00407423684373498, + 0.00874096155166626, + 0.00593697652220726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.009600185789167881 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003781944513321 + ], + "opposite_logprob_tokens": [ + -2.695037841796875 + ], + "logprob_reward": -0.07003781944513321, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003781944513321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676546186208725, + 0.1310884952545166, + 0.3526320159435272, + 0.028615077957510948, + 0.0340169295668602, + 0.09939295053482056, + 0.024554457515478134, + 0.06214051693677902, + 0.059256818145513535, + 0.040648117661476135 + ], + "credit_map": [ + 0.01709028147161007, + 0.017988841980695724, + 0.020180631428956985, + 0.018905825912952423, + 0.013266798108816147, + 0.045334696769714355, + 0.01485344860702753, + 0.02003411017358303, + 0.0141685726121068, + 0.01756243221461773, + 0.01140639465302229, + 0.038373496383428574, + 0.008004911243915558, + 0.019258107990026474, + 0.011013663373887539, + 0.011300913989543915, + 0.034470658749341965, + 0.054685551673173904, + 0.043194085359573364, + 0.048450667411088943, + 0.035666532814502716, + 0.03354160115122795, + 0.06131405383348465, + 0.04130886495113373, + 0.0031160989310592413, + 0.0024952758103609085, + 0.003364730393514037, + 0.0024972320534288883, + 0.0067962524481117725, + 0.002930270042270422, + 0.004354279488325119, + 0.0030609415844082832, + 0.0025875833816826344, + 0.004880339838564396, + 0.00495856674388051, + 0.004978272132575512, + 0.003212172770872712, + 0.004678085446357727, + 0.004136247560381889, + 0.004585668444633484, + 0.005830445792526007, + 0.014723619446158409, + 0.01502096839249134, + 0.014015079475939274, + 0.013485514558851719, + 0.013523481786251068, + 0.010993565432727337, + 0.011800281703472137, + 0.0019373840186744928, + 0.0028231199830770493, + 0.0030261969659477472, + 0.0032943624537438154, + 0.0033270956482738256, + 0.0034860281739383936, + 0.0024839933030307293, + 0.0041762785986065865, + 0.0038373712450265884, + 0.006453398615121841, + 0.005482208449393511, + 0.014771665446460247, + 0.005304334219545126, + 0.004557011183351278, + 0.012170895002782345, + 0.009563635103404522, + 0.0037985837552696466, + 0.006004706956446171, + 0.005507333669811487, + 0.007246694527566433, + 0.010448700748383999, + 0.006205561570823193, + 0.01314271055161953, + 0.006902528461068869, + 0.004792149178683758, + 0.005549189634621143, + 0.0061871446669101715, + 0.005544806364923716, + 0.004125102888792753, + 0.003092421218752861, + 0.005645927507430315, + 0.005711378995329142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07003781944513321 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546537443995476 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546537443995476, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546537443995476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12003987282514572, + 0.5162411332130432, + 0.02289753034710884, + 0.03699565306305885, + 0.011816958896815777, + 0.04047217220067978, + 0.01889919862151146, + 0.03773471340537071, + 0.08335664868354797, + 0.11154606938362122 + ], + "credit_map": [ + 0.014399268664419651, + 0.011867006309330463, + 0.014431129209697247, + 0.013780112378299236, + 0.010376924648880959, + 0.03015027567744255, + 0.00979090016335249, + 0.015244264155626297, + 0.07892756909132004, + 0.06490030884742737, + 0.04761141538619995, + 0.15857410430908203, + 0.026320265606045723, + 0.05213763937354088, + 0.046096403151750565, + 0.041673459112644196, + 0.0031304897274821997, + 0.0043410672806203365, + 0.0024098255671560764, + 0.002282300963997841, + 0.0022990175057202578, + 0.0020000396762043238, + 0.0030847263988107443, + 0.0033500625286251307, + 0.004187879152595997, + 0.003601722652092576, + 0.005328445695340633, + 0.00428337138146162, + 0.004485874902456999, + 0.005953318439424038, + 0.0038590081967413425, + 0.0052960338070988655, + 0.001310107414610684, + 0.001396781881339848, + 0.002078956924378872, + 0.0017391416477039456, + 0.0014320781920105219, + 0.0012977912556380033, + 0.0014863826800137758, + 0.0010757195996120572, + 0.0032182708382606506, + 0.006547587923705578, + 0.007241747807711363, + 0.00406468054279685, + 0.005128238815814257, + 0.0059142643585801125, + 0.004553289618343115, + 0.00380409206263721, + 0.0017130594933405519, + 0.002410031156614423, + 0.003705201903358102, + 0.002351114759221673, + 0.001545037142932415, + 0.0028440426103770733, + 0.0018057969864457846, + 0.0025249149184674025, + 0.0024975701235234737, + 0.0027092117816209793, + 0.002402701647952199, + 0.005013587418943644, + 0.00285445898771286, + 0.008138417266309261, + 0.005500138737261295, + 0.008618625812232494, + 0.010620523244142532, + 0.010689910501241684, + 0.010818139649927616, + 0.007871354930102825, + 0.013845276087522507, + 0.01117110438644886, + 0.009610085748136044, + 0.008730261586606503, + 0.01868966408073902, + 0.011048738844692707, + 0.02672513574361801, + 0.013226027600467205, + 0.0070628756657242775, + 0.008488303050398827, + 0.012604319490492344, + 0.013701018877327442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.012546537443995476 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054958924651145935 + ], + "opposite_logprob_tokens": [ + -2.9299588203430176 + ], + "logprob_reward": -0.054958924651145935, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958924651145935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1590176224708557, + 0.2461138516664505, + 0.06744235754013062, + 0.07682958990335464, + 0.05243920162320137, + 0.08188297599554062, + 0.10415022075176239, + 0.07533847540616989, + 0.05109100416302681, + 0.08569469302892685 + ], + "credit_map": [ + 0.019774489104747772, + 0.016031725332140923, + 0.020294900983572006, + 0.01640881411731243, + 0.013141498900949955, + 0.04127304628491402, + 0.011076980270445347, + 0.021016167476773262, + 0.03139256685972214, + 0.03108370490372181, + 0.024998405948281288, + 0.06932298094034195, + 0.01542678102850914, + 0.020690979436039925, + 0.02242751233279705, + 0.03077091835439205, + 0.010787148959934711, + 0.011747688055038452, + 0.008875422179698944, + 0.008190878666937351, + 0.004332961048930883, + 0.00623652758076787, + 0.008909430354833603, + 0.008362303487956524, + 0.010760681703686714, + 0.007956599816679955, + 0.00790299754589796, + 0.007410407532006502, + 0.015700625255703926, + 0.006894349120557308, + 0.011336513794958591, + 0.008867411874234676, + 0.005180640611797571, + 0.005523275583982468, + 0.006770786829292774, + 0.007956861518323421, + 0.005640451330691576, + 0.006942774169147015, + 0.006991118658334017, + 0.007433297578245401, + 0.007364929653704166, + 0.007106340955942869, + 0.013814994134008884, + 0.009361570701003075, + 0.012895113788545132, + 0.011204000562429428, + 0.012029902078211308, + 0.008106122724711895, + 0.007547114975750446, + 0.010215247981250286, + 0.02123318612575531, + 0.00996254663914442, + 0.01149049960076809, + 0.01588716171681881, + 0.008811897598206997, + 0.01900256983935833, + 0.004879874177277088, + 0.005331708583980799, + 0.006861381232738495, + 0.007991977035999298, + 0.0053193760104477406, + 0.018757624551653862, + 0.006037753541022539, + 0.020158778876066208, + 0.005815048702061176, + 0.0057630822993814945, + 0.0049287606962025166, + 0.0056311083026230335, + 0.012335801497101784, + 0.00451757712289691, + 0.006788182072341442, + 0.0053114453330636024, + 0.010575953871011734, + 0.009584257379174232, + 0.012291109189391136, + 0.011911521665751934, + 0.01184623409062624, + 0.006753916386514902, + 0.011262415908277035, + 0.011469284072518349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.054958924651145935 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043059516698122025 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.043059516698122025, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059516698122025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11106383055448532, + 0.0931255891919136, + 0.04200081527233124, + 0.08128693699836731, + 0.07556332647800446, + 0.19429919123649597, + 0.05691252276301384, + 0.058010369539260864, + 0.03752398118376732, + 0.2502134144306183 + ], + "credit_map": [ + 0.01825021021068096, + 0.014754275791347027, + 0.010786829516291618, + 0.01650153659284115, + 0.01104861032217741, + 0.01857810840010643, + 0.010015688836574554, + 0.011128569021821022, + 0.00948997400701046, + 0.01058789063245058, + 0.011210381053388119, + 0.01164297852665186, + 0.013348236680030823, + 0.010357209481298923, + 0.012648687697947025, + 0.013840226456522942, + 0.004916596692055464, + 0.005708915647119284, + 0.005228858906775713, + 0.004461446311324835, + 0.0038733966648578644, + 0.004559552762657404, + 0.006012353114783764, + 0.00723969517275691, + 0.007903806865215302, + 0.007240045350044966, + 0.006370732095092535, + 0.007938243448734283, + 0.019611243158578873, + 0.009426124393939972, + 0.011176246218383312, + 0.011620501056313515, + 0.009346174076199532, + 0.0073633999563753605, + 0.017466159537434578, + 0.008606676943600178, + 0.0052862269803881645, + 0.009974614717066288, + 0.008372403681278229, + 0.009147671051323414, + 0.008771369233727455, + 0.015589363873004913, + 0.01559802982956171, + 0.026283161714673042, + 0.03700557351112366, + 0.02835727296769619, + 0.04409468173980713, + 0.018599729984998703, + 0.005246572662144899, + 0.005284832324832678, + 0.008258141577243805, + 0.0068777878768742085, + 0.0051816729828715324, + 0.007608131971210241, + 0.0074179330840706825, + 0.011037454009056091, + 0.006178529933094978, + 0.005779688712209463, + 0.008585942909121513, + 0.010210886597633362, + 0.006016901694238186, + 0.0038324608467519283, + 0.008541000075638294, + 0.008864959701895714, + 0.004234099294990301, + 0.004504132084548473, + 0.004551981575787067, + 0.005011132452636957, + 0.007567423861473799, + 0.0030683069489896297, + 0.005322516430169344, + 0.0032643890008330345, + 0.017366493120789528, + 0.025774294510483742, + 0.023483507335186005, + 0.03602368384599686, + 0.03164711222052574, + 0.022042086347937584, + 0.05463968217372894, + 0.03923657163977623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.043059516698122025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up, shot with a sleek, doll-like backward tracking angle, perfectly captures the determined face of an East Asian woman in her twenties as she steadily moves forward. Her brow is slightly furrowed in concentration, her dark, focused eyes fixed on a single point off-screen, her lips pressed tightly together. Her sleek, shoulder-length black hair, slightly damp from the cool night air, captures the ever-changing ambient light. She wears a charcoal gray turtleneck windbreaker, its waterproof fabric reflecting a kaleidoscopic hue. The nighttime cityscape behind her becomes a vibrant, abstract tapestry composed of motion-blurred streaks—electric blue, neon pink, and warm yellow from traffic and storefronts—with fleeting highlights traced across her sharp cheekbones and jawline. Her purposeful walk is relentless and steady, her head held level, creating a powerful sense of unwavering determination and forward momentum amid the chaotic, out-of-focus beauty of the urban environment, establishing an atmosphere of intense loneliness.", + "seed": 27852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_6621_seed_27852_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": -0.036374611966311934, + "vlm_logprob": -0.036374611966311934, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.036374611966311934, + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012016643770039082 + ], + "opposite_logprob_tokens": [ + -4.637016773223877 + ], + "logprob_reward": -0.012016643770039082, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012016643770039082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270550936460495, + 0.04864267259836197, + 0.036089710891246796, + 0.08520399034023285, + 0.062212031334638596, + 0.05146754905581474, + 0.14264652132987976, + 0.13304390013217926, + 0.07866977900266647, + 0.09147286415100098 + ], + "credit_map": [ + 0.021808788180351257, + 0.05400576442480087, + 0.03607173636555672, + 0.017657577991485596, + 0.04524678364396095, + 0.01155982818454504, + 0.02912995219230652, + 0.055070534348487854, + 0.0044683837331831455, + 0.009758087806403637, + 0.005272383335977793, + 0.005312809720635414, + 0.008691591210663319, + 0.0051494017243385315, + 0.005048580002039671, + 0.004941439721733332, + 0.0034108508843928576, + 0.0030641593039035797, + 0.004241084214299917, + 0.0027942457236349583, + 0.00726691959425807, + 0.005167740397155285, + 0.0037473919801414013, + 0.00639732601121068, + 0.0061697824858129025, + 0.010174499824643135, + 0.003391236998140812, + 0.00907391868531704, + 0.013710465282201767, + 0.01069315243512392, + 0.018350377678871155, + 0.013640566729009151, + 0.006630693096667528, + 0.011612356640398502, + 0.00791139155626297, + 0.005384623073041439, + 0.008152014575898647, + 0.0030688107945024967, + 0.007950833067297935, + 0.011501316912472248, + 0.003573607886210084, + 0.003173565724864602, + 0.009110977873206139, + 0.00427710497751832, + 0.004397472832351923, + 0.0024588170927017927, + 0.007783886045217514, + 0.01669212244451046, + 0.008314303122460842, + 0.010653185658156872, + 0.017464561387896538, + 0.018249988555908203, + 0.027605069801211357, + 0.007373311091214418, + 0.027363283559679985, + 0.025622839108109474, + 0.016945013776421547, + 0.013486811891198158, + 0.01613587886095047, + 0.017431724816560745, + 0.015170102939009666, + 0.018566688522696495, + 0.01728343777358532, + 0.01802426017820835, + 0.014071987941861153, + 0.007900422438979149, + 0.008663651533424854, + 0.014480087906122208, + 0.007281734142452478, + 0.007468503899872303, + 0.008865820243954659, + 0.009937579743564129, + 0.006455839611589909, + 0.012430978007614613, + 0.005919557064771652, + 0.013744207099080086, + 0.016327960416674614, + 0.004978632088750601, + 0.019067510962486267, + 0.012548188678920269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.012016643770039082 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207192316651344 + ], + "opposite_logprob_tokens": [ + -2.8120720386505127 + ], + "logprob_reward": -0.06207192316651344, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207192316651344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19475308060646057, + 0.06538339704275131, + 0.022530585527420044, + 0.05562268942594528, + 0.082563616335392, + 0.019989825785160065, + 0.032701943069696426, + 0.09299030154943466, + 0.11789170652627945, + 0.3155728578567505 + ], + "credit_map": [ + 0.011249547824263573, + 0.06271281838417053, + 0.01580372266471386, + 0.008074677549302578, + 0.017972098663449287, + 0.010269206017255783, + 0.014219855889678001, + 0.05445116013288498, + 0.010658920742571354, + 0.005896859336644411, + 0.0075356075540184975, + 0.007054320070892572, + 0.0096433749422431, + 0.006572797428816557, + 0.008893736638128757, + 0.009127779863774776, + 0.004543357994407415, + 0.0022715902887284756, + 0.0028820603620260954, + 0.002352065173909068, + 0.003834919771179557, + 0.0017185219330713153, + 0.002622980158776045, + 0.0023050897289067507, + 0.005436834879219532, + 0.006518679205328226, + 0.003978079650551081, + 0.0055819544941186905, + 0.006616534199565649, + 0.008168266154825687, + 0.011361878365278244, + 0.00796046108007431, + 0.011309637688100338, + 0.015074742026627064, + 0.009815684519708157, + 0.0077039338648319244, + 0.010724339634180069, + 0.0046289958991110325, + 0.00967367086559534, + 0.013632616959512234, + 0.0016739574493840337, + 0.0018538690637797117, + 0.002555213402956724, + 0.005053830333054066, + 0.0031797955743968487, + 0.0007305433391593397, + 0.002036615740507841, + 0.00290600280277431, + 0.002311292802914977, + 0.003005895996466279, + 0.00306961452588439, + 0.0038388315588235855, + 0.008673068135976791, + 0.0016489285044372082, + 0.005916744004935026, + 0.004237567540258169, + 0.009688135236501694, + 0.0069249216467142105, + 0.010471636429429054, + 0.014336206018924713, + 0.010624205693602562, + 0.01010836474597454, + 0.019427256658673286, + 0.011409575119614601, + 0.0062725478783249855, + 0.014994053170084953, + 0.010932530276477337, + 0.036186955869197845, + 0.014349300414323807, + 0.0049302950501441956, + 0.023753676563501358, + 0.006472339387983084, + 0.02258658967912197, + 0.04300336167216301, + 0.01839006133377552, + 0.0436716303229332, + 0.05855116993188858, + 0.015724247321486473, + 0.07005802541971207, + 0.04358777031302452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.06207192316651344 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19936315715312958, + 0.09694567322731018, + 0.027640080079436302, + 0.3313336670398712, + 0.01837879605591297, + 0.0223004799336195, + 0.011133285239338875, + 0.23594360053539276, + 0.013272132724523544, + 0.043689142912626266 + ], + "credit_map": [ + 0.011198028922080994, + 0.06568867713212967, + 0.01900048740208149, + 0.008493623696267605, + 0.014596868306398392, + 0.010209381580352783, + 0.012869545258581638, + 0.057306546717882156, + 0.011130557395517826, + 0.00852880161255598, + 0.012523343786597252, + 0.007106941659003496, + 0.009661823511123657, + 0.021839629858732224, + 0.014350064098834991, + 0.011804507113993168, + 0.003670369740575552, + 0.002851860597729683, + 0.0042671095579862595, + 0.002644266001880169, + 0.003486362751573324, + 0.004707451909780502, + 0.0025830762460827827, + 0.0034295832738280296, + 0.024756159633398056, + 0.04086816683411598, + 0.01667356677353382, + 0.03254857286810875, + 0.04728742316365242, + 0.052312497049570084, + 0.06569952517747879, + 0.05118779465556145, + 0.0025297857355326414, + 0.0043199509382247925, + 0.002005025278776884, + 0.0018016370013356209, + 0.0026620528660714626, + 0.0010115590412169695, + 0.002040290739387274, + 0.002008493058383465, + 0.0025677415542304516, + 0.00184765481390059, + 0.0038277858402580023, + 0.0018988567171618342, + 0.0025416298303753138, + 0.002186874160543084, + 0.0025461260229349136, + 0.004883811809122562, + 0.0012296081986278296, + 0.0012283696560189128, + 0.0016098173800855875, + 0.0011817237827926874, + 0.001778436591848731, + 0.000726719677913934, + 0.001980556407943368, + 0.0013980536023154855, + 0.02886638604104519, + 0.021419094875454903, + 0.02882266230881214, + 0.02982725389301777, + 0.024272797629237175, + 0.030309783294796944, + 0.03984870761632919, + 0.03257690742611885, + 0.0019527244148775935, + 0.0020387887489050627, + 0.0010391368996351957, + 0.002396585186943412, + 0.0013557261554524302, + 0.0008994992822408676, + 0.002709988271817565, + 0.000879684288520366, + 0.0027316007763147354, + 0.005007628351449966, + 0.0033475589007139206, + 0.007127209100872278, + 0.0075097279623150826, + 0.002985341241583228, + 0.007835239171981812, + 0.007144838105887175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.007633799687027931 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205063685774803 + ], + "opposite_logprob_tokens": [ + -2.8120505809783936 + ], + "logprob_reward": -0.06205063685774803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205063685774803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31428393721580505, + 0.0562407411634922, + 0.05956600233912468, + 0.1036769449710846, + 0.052087731659412384, + 0.04019290581345558, + 0.06366793066263199, + 0.07862931489944458, + 0.1350650042295456, + 0.09658949077129364 + ], + "credit_map": [ + 0.019453471526503563, + 0.09595515578985214, + 0.03018869459629059, + 0.014585589990019798, + 0.029688257724046707, + 0.012869050726294518, + 0.01901891827583313, + 0.09252478182315826, + 0.011116528883576393, + 0.007351364474743605, + 0.007552755065262318, + 0.005861710757017136, + 0.006391441449522972, + 0.005503205116838217, + 0.005123880226165056, + 0.007339856121689081, + 0.01319371908903122, + 0.004124816041439772, + 0.012559051625430584, + 0.004298614338040352, + 0.0051694270223379135, + 0.004908295813947916, + 0.007720750756561756, + 0.007591329514980316, + 0.01652752049267292, + 0.00843543466180563, + 0.010742821730673313, + 0.007661233656108379, + 0.01216831337660551, + 0.005897377617657185, + 0.024564914405345917, + 0.017679324373602867, + 0.007130085490643978, + 0.011861424893140793, + 0.0041692801751196384, + 0.005613830406218767, + 0.006727189756929874, + 0.0027680937200784683, + 0.005980943329632282, + 0.007836887612938881, + 0.002060956321656704, + 0.0024436484090983868, + 0.007795528043061495, + 0.0034812670201063156, + 0.0061413198709487915, + 0.0034674585331231356, + 0.004462614189833403, + 0.010340111330151558, + 0.0065856254659593105, + 0.006010161247104406, + 0.01259492989629507, + 0.005138808395713568, + 0.007817087695002556, + 0.0037601098883897066, + 0.009453386068344116, + 0.012307820841670036, + 0.010508869774639606, + 0.005776438862085342, + 0.008414041250944138, + 0.008406471461057663, + 0.007710120640695095, + 0.008993569761514664, + 0.013572361320257187, + 0.015247435308992863, + 0.026878977194428444, + 0.01149889174848795, + 0.015561566688120365, + 0.03197351098060608, + 0.007427833508700132, + 0.01031606923788786, + 0.01756402850151062, + 0.013844115659594536, + 0.006309139542281628, + 0.013911016285419464, + 0.006217853166162968, + 0.015330999158322811, + 0.02064151130616665, + 0.004943811800330877, + 0.016103170812129974, + 0.013131992891430855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.06205063685774803 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810005635023117 + ], + "opposite_logprob_tokens": [ + -3.288100004196167 + ], + "logprob_reward": -0.03810005635023117, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810005635023117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1282983422279358, + 0.09425336122512817, + 0.13918982446193695, + 0.03204667568206787, + 0.16414812207221985, + 0.019350139424204826, + 0.06310474872589111, + 0.24243681132793427, + 0.03365439549088478, + 0.08351751416921616 + ], + "credit_map": [ + 0.00920828990638256, + 0.02251295931637287, + 0.013729974627494812, + 0.010641638189554214, + 0.021209342405200005, + 0.0072480784729123116, + 0.01360019389539957, + 0.030147874727845192, + 0.011310000903904438, + 0.01865888200700283, + 0.013056394644081593, + 0.012441317550837994, + 0.014293747022747993, + 0.00577097712084651, + 0.008595837280154228, + 0.010126215405762196, + 0.018414722755551338, + 0.01010852586477995, + 0.02163350023329258, + 0.01024079229682684, + 0.015516567975282669, + 0.01996171846985817, + 0.013394886627793312, + 0.029919130727648735, + 0.0038541259709745646, + 0.0031470204703509808, + 0.003749396186321974, + 0.0037248050794005394, + 0.004842641297727823, + 0.003595098853111267, + 0.004102622624486685, + 0.005030967760831118, + 0.023165633901953697, + 0.03872004151344299, + 0.013427061960101128, + 0.018934179097414017, + 0.021184680983424187, + 0.007040421478450298, + 0.01950971782207489, + 0.02216639742255211, + 0.0015999043826013803, + 0.0017261393368244171, + 0.0032939233351498842, + 0.002198980189859867, + 0.002720335964113474, + 0.0019256641389802098, + 0.002705135615542531, + 0.003180058440193534, + 0.007218687795102596, + 0.0071379574947059155, + 0.009285081177949905, + 0.005955603439360857, + 0.0079713249579072, + 0.006212614942342043, + 0.010964780114591122, + 0.00835870485752821, + 0.028055528178811073, + 0.019372867420315742, + 0.028704460710287094, + 0.030339645221829414, + 0.02656780369579792, + 0.02757551148533821, + 0.050940707325935364, + 0.030880311504006386, + 0.003312300890684128, + 0.00451227743178606, + 0.0035584266297519207, + 0.006952650845050812, + 0.003812082577496767, + 0.0022209188900887966, + 0.005315801594406366, + 0.003969939891248941, + 0.006978503428399563, + 0.012534957379102707, + 0.0047040036879479885, + 0.011680008843541145, + 0.0165242999792099, + 0.003783999476581812, + 0.01729615032672882, + 0.010015593841671944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.03810005635023117 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up, shot with a sleek, doll-like backward tracking angle, perfectly captures the determined face of an East Asian woman in her twenties as she steadily moves forward. Her brow is slightly furrowed in concentration, her dark, focused eyes fixed on a single point off-screen, her lips pressed tightly together. Her sleek, shoulder-length black hair, slightly damp from the cool night air, captures the ever-changing ambient light. She wears a charcoal gray turtleneck windbreaker, its waterproof fabric reflecting a kaleidoscopic hue. The nighttime cityscape behind her becomes a vibrant, abstract tapestry composed of motion-blurred streaks—electric blue, neon pink, and warm yellow from traffic and storefronts—with fleeting highlights traced across her sharp cheekbones and jawline. Her purposeful walk is relentless and steady, her head held level, creating a powerful sense of unwavering determination and forward momentum amid the chaotic, out-of-focus beauty of the urban environment, establishing an atmosphere of intense loneliness.", + "seed": 27853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_6621_seed_27853_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": -0.02392180860042572, + "vlm_logprob": -0.02392180860042572, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02392180860042572, + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007481177803128958 + ], + "opposite_logprob_tokens": [ + -5.257481098175049 + ], + "logprob_reward": -0.007481177803128958, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007481177803128958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2067907601594925, + 0.07879503816366196, + 0.05836894363164902, + 0.049062199890613556, + 0.1175418347120285, + 0.11289383471012115, + 0.05515899881720543, + 0.10633485019207001, + 0.13745999336242676, + 0.07759350538253784 + ], + "credit_map": [ + 0.019801493734121323, + 0.01594102568924427, + 0.03396248817443848, + 0.022179467603564262, + 0.025742897763848305, + 0.027568751946091652, + 0.020417703315615654, + 0.041176922619342804, + 0.009749779477715492, + 0.010344413109123707, + 0.008296522311866283, + 0.007491007912904024, + 0.009681235998868942, + 0.009437966160476208, + 0.013458406552672386, + 0.010335719212889671, + 0.006832244340330362, + 0.008235041983425617, + 0.0068037318997085094, + 0.004881612490862608, + 0.0084012346342206, + 0.005765125621110201, + 0.009887329302728176, + 0.007562627084553242, + 0.005373700056225061, + 0.006993040442466736, + 0.0032128598541021347, + 0.005209730938076973, + 0.00548851303756237, + 0.005947093479335308, + 0.006851985119283199, + 0.009985281154513359, + 0.013196049258112907, + 0.017477547749876976, + 0.011653498746454716, + 0.012746117077767849, + 0.023746592923998833, + 0.010611358098685741, + 0.0199597105383873, + 0.008150966838002205, + 0.01423117145895958, + 0.012687272392213345, + 0.013946929015219212, + 0.016452210023999214, + 0.025604192167520523, + 0.006704369094222784, + 0.015333310700953007, + 0.0079343905672431, + 0.0050403885543346405, + 0.011668209917843342, + 0.007081399206072092, + 0.005215003155171871, + 0.005401160567998886, + 0.005931066814810038, + 0.008109583519399166, + 0.006712192669510841, + 0.007421605754643679, + 0.011105154640972614, + 0.008590169250965118, + 0.011555002070963383, + 0.031241342425346375, + 0.0060298871248960495, + 0.017371682450175285, + 0.013020016252994537, + 0.029576003551483154, + 0.007900604978203773, + 0.01242729090154171, + 0.00905479397624731, + 0.017717912793159485, + 0.013319363817572594, + 0.021407993510365486, + 0.026056041941046715, + 0.005981030408293009, + 0.006694048177450895, + 0.007020600605756044, + 0.006788103841245174, + 0.015991833060979843, + 0.005804580636322498, + 0.016990510746836662, + 0.012322812341153622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.007481177803128958 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430825836956501 + ], + "opposite_logprob_tokens": [ + -3.1680824756622314 + ], + "logprob_reward": -0.0430825836956501, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430825836956501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16718827188014984, + 0.04891242831945419, + 0.030433887615799904, + 0.025224149227142334, + 0.17205092310905457, + 0.3560152053833008, + 0.0352913923561573, + 0.09935790300369263, + 0.023747121915221214, + 0.04177869111299515 + ], + "credit_map": [ + 0.022178715094923973, + 0.009611514396965504, + 0.04004771634936333, + 0.012855676002800465, + 0.010917840525507927, + 0.02383360080420971, + 0.011193377897143364, + 0.03654983267188072, + 0.006655889097601175, + 0.003270563203841448, + 0.002481425879523158, + 0.00419262470677495, + 0.0069318488240242004, + 0.005857391748577356, + 0.005982798989862204, + 0.013539884239435196, + 0.0029728796798735857, + 0.0039846766740083694, + 0.004468854051083326, + 0.004435625392943621, + 0.005007104482501745, + 0.003121706424281001, + 0.0035066239070147276, + 0.0029364172369241714, + 0.0024660176131874323, + 0.00408735079690814, + 0.0013849418610334396, + 0.0026429651770740747, + 0.003947147633880377, + 0.0023418059572577477, + 0.0038549485616385937, + 0.004498971626162529, + 0.012891262769699097, + 0.03069855459034443, + 0.012842193245887756, + 0.02255590260028839, + 0.03690658137202263, + 0.015637898817658424, + 0.022706128656864166, + 0.01781240478157997, + 0.0571909137070179, + 0.038765788078308105, + 0.06052032485604286, + 0.05036419630050659, + 0.04005607217550278, + 0.02969786524772644, + 0.05117001011967659, + 0.02825002931058407, + 0.0037295527290552855, + 0.0056680189445614815, + 0.006340949330478907, + 0.004596624989062548, + 0.004769073333591223, + 0.0025289535988122225, + 0.005299786105751991, + 0.0023584316950291395, + 0.0072592017240822315, + 0.01155316922813654, + 0.008828796446323395, + 0.010783315636217594, + 0.026341687887907028, + 0.0067718857899308205, + 0.014256471768021584, + 0.013563381507992744, + 0.004494148772209883, + 0.0013218754902482033, + 0.0019260209519416094, + 0.0015302004758268595, + 0.003060593269765377, + 0.0029582034330815077, + 0.004113018047064543, + 0.00434306263923645, + 0.0031482053454965353, + 0.004154855851083994, + 0.0033341834787279367, + 0.006039402913302183, + 0.008223693817853928, + 0.0021743588149547577, + 0.00893973745405674, + 0.005764253903180361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0430825836956501 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022055604495108128 + ], + "opposite_logprob_tokens": [ + -6.1272053718566895 + ], + "logprob_reward": -0.0022055604495108128, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022055604495108128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22311630845069885, + 0.06742911785840988, + 0.06389540433883667, + 0.07371903210878372, + 0.14167821407318115, + 0.01799073815345764, + 0.03964471071958542, + 0.2153654247522354, + 0.030345510691404343, + 0.1268155425786972 + ], + "credit_map": [ + 0.030184902250766754, + 0.016584796831011772, + 0.05379842221736908, + 0.021041937172412872, + 0.017781782895326614, + 0.029880719259381294, + 0.014663473702967167, + 0.03918028250336647, + 0.009772920981049538, + 0.01035027764737606, + 0.009167313575744629, + 0.006689742673188448, + 0.007594467140734196, + 0.008635026402771473, + 0.008897433057427406, + 0.006321933586150408, + 0.0078698405995965, + 0.007437582593411207, + 0.013024627231061459, + 0.006231965962797403, + 0.00786010455340147, + 0.006283331662416458, + 0.008200911805033684, + 0.006987037602812052, + 0.008667702786624432, + 0.011125774122774601, + 0.007211262825876474, + 0.005454297177493572, + 0.008093182928860188, + 0.010402163490653038, + 0.010843164287507534, + 0.011921477504074574, + 0.016233505681157112, + 0.02173343300819397, + 0.012052576057612896, + 0.015580803155899048, + 0.028056394308805466, + 0.014058158732950687, + 0.02079450525343418, + 0.013168844394385815, + 0.0015961487079039216, + 0.0024496109690517187, + 0.0015501618618145585, + 0.0022273005452007055, + 0.003946571610867977, + 0.0016592503525316715, + 0.0033642861526459455, + 0.0011974069057032466, + 0.0037088212557137012, + 0.008338750340044498, + 0.006105287931859493, + 0.005322138778865337, + 0.0036443741992115974, + 0.004440160468220711, + 0.004556015133857727, + 0.0035291591193526983, + 0.014391651377081871, + 0.022421739995479584, + 0.016408288851380348, + 0.02453898824751377, + 0.07085197418928146, + 0.012286735698580742, + 0.03724781423807144, + 0.01721823215484619, + 0.0029639939311891794, + 0.0018104148330166936, + 0.004573678132146597, + 0.003758247010409832, + 0.004249852616339922, + 0.0032326665241271257, + 0.005081386771053076, + 0.00467526912689209, + 0.011681263335049152, + 0.014457684010267258, + 0.01362683903425932, + 0.01256473083049059, + 0.029467452317476273, + 0.012926694937050343, + 0.016526861116290092, + 0.015564017929136753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0022055604495108128 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048651713877916336 + ], + "opposite_logprob_tokens": [ + -3.048651695251465 + ], + "logprob_reward": -0.048651713877916336, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048651713877916336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3872542381286621, + 0.07653561234474182, + 0.0799115002155304, + 0.13538697361946106, + 0.03235345706343651, + 0.057002462446689606, + 0.08729694038629532, + 0.057835981249809265, + 0.04615876451134682, + 0.040264103561639786 + ], + "credit_map": [ + 0.055204201489686966, + 0.01289943978190422, + 0.10373096913099289, + 0.01752190850675106, + 0.024128101766109467, + 0.05839819088578224, + 0.020852327346801758, + 0.09451906383037567, + 0.006293740123510361, + 0.007270808331668377, + 0.005665754433721304, + 0.006469618529081345, + 0.009007584303617477, + 0.00960059929639101, + 0.014198892749845982, + 0.018028616905212402, + 0.009358637034893036, + 0.010121048428118229, + 0.010513938963413239, + 0.008958461694419384, + 0.011419261805713177, + 0.009266667068004608, + 0.011846702545881271, + 0.008426781743764877, + 0.014196684584021568, + 0.025951040908694267, + 0.008595614694058895, + 0.016338950023055077, + 0.018260525539517403, + 0.013807226903736591, + 0.020005352795124054, + 0.018231574445962906, + 0.004218300338834524, + 0.004278053063899279, + 0.0026261769235134125, + 0.004051098600029945, + 0.007852262817323208, + 0.0026777430903166533, + 0.004566633608192205, + 0.0020831904839724302, + 0.006849124561995268, + 0.006086172070354223, + 0.009403492324054241, + 0.0071334922686219215, + 0.008101769722998142, + 0.006635831203311682, + 0.008686554618179798, + 0.004106022417545319, + 0.01018240861594677, + 0.01872408390045166, + 0.014373705722391605, + 0.011648442596197128, + 0.008863639086484909, + 0.007339212577790022, + 0.012183904647827148, + 0.003981538116931915, + 0.0033461246639490128, + 0.005094089545309544, + 0.003571908688172698, + 0.007358994800597429, + 0.016811585053801537, + 0.003150696400552988, + 0.009451017715036869, + 0.009051565080881119, + 0.004120984114706516, + 0.0031921074260026217, + 0.009713554754853249, + 0.006115708965808153, + 0.006236029788851738, + 0.004471227526664734, + 0.005974357947707176, + 0.006334794219583273, + 0.003172316588461399, + 0.0037080671172589064, + 0.005408056546002626, + 0.003619173541665077, + 0.009062070399522781, + 0.0025225027929991484, + 0.007630627136677504, + 0.005141291301697493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.048651713877916336 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188007175922394 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188007175922394, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188007175922394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39784175157546997, + 0.07564026862382889, + 0.05864192172884941, + 0.04890331253409386, + 0.055884454399347305, + 0.04494128003716469, + 0.1258772909641266, + 0.03776773065328598, + 0.09847886115312576, + 0.05602322891354561 + ], + "credit_map": [ + 0.05368686839938164, + 0.012937680818140507, + 0.10197200626134872, + 0.01650412566959858, + 0.024194758385419846, + 0.06544659286737442, + 0.022731972858309746, + 0.10036775469779968, + 0.008125478401780128, + 0.005160324275493622, + 0.008389601483941078, + 0.00600465526804328, + 0.012982316315174103, + 0.008847564458847046, + 0.009956779889762402, + 0.016173550859093666, + 0.006555491592735052, + 0.010727159678936005, + 0.005769150331616402, + 0.008181910030543804, + 0.007309827022254467, + 0.006737290415912867, + 0.006576961372047663, + 0.006784128490835428, + 0.003958961460739374, + 0.010405690409243107, + 0.003111058147624135, + 0.006461330223828554, + 0.00738545460626483, + 0.004571774508804083, + 0.007395632565021515, + 0.005613412708044052, + 0.005199756007641554, + 0.008952961303293705, + 0.004957713186740875, + 0.006811989471316338, + 0.010713767260313034, + 0.005141241475939751, + 0.008836757391691208, + 0.00527026504278183, + 0.0036164738703519106, + 0.005144866649061441, + 0.0032847768161445856, + 0.007082933560013771, + 0.011734037660062313, + 0.0037699101958423853, + 0.0065675643272697926, + 0.003740717889741063, + 0.011382990516722202, + 0.02563621662557125, + 0.01752510666847229, + 0.013788744807243347, + 0.010972774587571621, + 0.016783269122242928, + 0.015110895037651062, + 0.014677291736006737, + 0.002231763442978263, + 0.0037969346158206463, + 0.003088278230279684, + 0.005941979121416807, + 0.0092873340472579, + 0.002672798465937376, + 0.005991646554321051, + 0.00475699407979846, + 0.010382500477135181, + 0.0066403537057340145, + 0.008681762032210827, + 0.011947454884648323, + 0.015018744394183159, + 0.01461445726454258, + 0.015186889097094536, + 0.016006706282496452, + 0.004631950985640287, + 0.005304699297994375, + 0.0063054910860955715, + 0.005123569630086422, + 0.013765737414360046, + 0.00382374320179224, + 0.00870999414473772, + 0.008358045481145382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.018188007175922394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle close-up, shot with a sleek, doll-like backward tracking angle, perfectly captures the determined face of an East Asian woman in her twenties as she steadily moves forward. Her brow is slightly furrowed in concentration, her dark, focused eyes fixed on a single point off-screen, her lips pressed tightly together. Her sleek, shoulder-length black hair, slightly damp from the cool night air, captures the ever-changing ambient light. She wears a charcoal gray turtleneck windbreaker, its waterproof fabric reflecting a kaleidoscopic hue. The nighttime cityscape behind her becomes a vibrant, abstract tapestry composed of motion-blurred streaks—electric blue, neon pink, and warm yellow from traffic and storefronts—with fleeting highlights traced across her sharp cheekbones and jawline. Her purposeful walk is relentless and steady, her head held level, creating a powerful sense of unwavering determination and forward momentum amid the chaotic, out-of-focus beauty of the urban environment, establishing an atmosphere of intense loneliness.", + "seed": 27854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_14/index_6621_seed_27854_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": -0.02562681301496923, + "vlm_logprob": -0.02562681301496923, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02562681301496923, + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006316459737718105 + ], + "opposite_logprob_tokens": [ + -5.506316661834717 + ], + "logprob_reward": -0.006316459737718105, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006316459737718105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23147214949131012, + 0.12122012674808502, + 0.08855685591697693, + 0.12370146065950394, + 0.04339886084198952, + 0.07011229544878006, + 0.048022978007793427, + 0.07914222776889801, + 0.06769543886184692, + 0.12667757272720337 + ], + "credit_map": [ + 0.00934484414756298, + 0.028550824150443077, + 0.02109067142009735, + 0.06456901878118515, + 0.03173335641622543, + 0.013459423556923866, + 0.04273127019405365, + 0.01999274455010891, + 0.007557584438472986, + 0.011477064341306686, + 0.018599912524223328, + 0.023848645389080048, + 0.007819479331374168, + 0.007436420302838087, + 0.01144229806959629, + 0.03303871676325798, + 0.0037812120281159878, + 0.019519776105880737, + 0.004700048826634884, + 0.007866577245295048, + 0.006715289782732725, + 0.0032788501121103764, + 0.029913192614912987, + 0.012781905941665173, + 0.007913275621831417, + 0.011248440481722355, + 0.006947399117052555, + 0.049303580075502396, + 0.0056215194053947926, + 0.0040464820340275764, + 0.015681838616728783, + 0.022938929498195648, + 0.005012099631130695, + 0.006466232240200043, + 0.005094141233712435, + 0.006224931217730045, + 0.004863090347498655, + 0.0028453858103603125, + 0.006668676622211933, + 0.006224303971976042, + 0.005260472651571035, + 0.010108078829944134, + 0.0139365429058671, + 0.011625244282186031, + 0.005584125407040119, + 0.006272135768085718, + 0.00735996151342988, + 0.009965733624994755, + 0.0039032534696161747, + 0.007363008335232735, + 0.007513502147048712, + 0.0075568207539618015, + 0.00354459835216403, + 0.005221443250775337, + 0.005656047258526087, + 0.007264305371791124, + 0.008664165623486042, + 0.009641521610319614, + 0.018825676292181015, + 0.008227763697504997, + 0.003843082347884774, + 0.006908163893967867, + 0.006173853296786547, + 0.016857998445630074, + 0.005662855692207813, + 0.008461209014058113, + 0.005502769723534584, + 0.007068784441798925, + 0.007860388606786728, + 0.010450981557369232, + 0.015748051926493645, + 0.006940399296581745, + 0.012731174007058144, + 0.02242237515747547, + 0.011426226235926151, + 0.020892437547445297, + 0.015933237969875336, + 0.009630446322262287, + 0.02392488904297352, + 0.009716792032122612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.006316459737718105 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062066324055194855 + ], + "opposite_logprob_tokens": [ + -2.8120663166046143 + ], + "logprob_reward": -0.062066324055194855, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062066324055194855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07965999096632004, + 0.08787934482097626, + 0.09233316779136658, + 0.3211829364299774, + 0.07139326632022858, + 0.0932496041059494, + 0.04025318846106529, + 0.04308011382818222, + 0.055281560868024826, + 0.11568687856197357 + ], + "credit_map": [ + 0.004799494054168463, + 0.014760985039174557, + 0.010108092799782753, + 0.00889445748180151, + 0.014219514094293118, + 0.005080668721348047, + 0.012371772900223732, + 0.00942500215023756, + 0.004923779517412186, + 0.007215022575110197, + 0.019295066595077515, + 0.009540756233036518, + 0.00903987605124712, + 0.003939826972782612, + 0.008213738910853863, + 0.02571127749979496, + 0.004448033403605223, + 0.024141842499375343, + 0.014330324716866016, + 0.011228974908590317, + 0.011684801429510117, + 0.003892096458002925, + 0.00698072137311101, + 0.015626369044184685, + 0.019547708332538605, + 0.03050345554947853, + 0.012288793921470642, + 0.1296781301498413, + 0.012568511068820953, + 0.010734908282756805, + 0.04088529571890831, + 0.06497611850500107, + 0.0051408978179097176, + 0.014686167240142822, + 0.010536876507103443, + 0.010855890810489655, + 0.0051818364299833775, + 0.002553635509684682, + 0.01309068500995636, + 0.0093472795560956, + 0.005608221981674433, + 0.015909794718027115, + 0.019153885543346405, + 0.014918056316673756, + 0.008704234845936298, + 0.004482808522880077, + 0.012403448112308979, + 0.01206915732473135, + 0.003960250411182642, + 0.005855500232428312, + 0.005182995460927486, + 0.007182525936514139, + 0.0036297107581049204, + 0.004064727108925581, + 0.004722691606730223, + 0.005654787179082632, + 0.00412760628387332, + 0.005595008842647076, + 0.008058887906372547, + 0.00464992132037878, + 0.003021403681486845, + 0.005417549982666969, + 0.00467947730794549, + 0.007530259899795055, + 0.005558573640882969, + 0.00612986646592617, + 0.005262198392301798, + 0.008524027653038502, + 0.007945265620946884, + 0.0036140300799161196, + 0.012807989493012428, + 0.005439613480120897, + 0.010182988829910755, + 0.03507741168141365, + 0.0076014334335923195, + 0.0104822413995862, + 0.009918060153722763, + 0.005408428609371185, + 0.01900680735707283, + 0.018009502440690994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.062066324055194855 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11701962351799011, + 0.07031102478504181, + 0.3735457956790924, + 0.08121241629123688, + 0.03833950683474541, + 0.0640951469540596, + 0.032654136419296265, + 0.027100780978798866, + 0.038445062935352325, + 0.15727652609348297 + ], + "credit_map": [ + 0.006379183381795883, + 0.011048443615436554, + 0.01165486965328455, + 0.0299088042229414, + 0.014940825290977955, + 0.005842262879014015, + 0.01815519854426384, + 0.01909004896879196, + 0.005821962375193834, + 0.007312806788831949, + 0.011700494214892387, + 0.01418758649379015, + 0.0057943095453083515, + 0.004471758846193552, + 0.006763585843145847, + 0.01425852905958891, + 0.009130677208304405, + 0.12078731507062912, + 0.010880555957555771, + 0.016436424106359482, + 0.018666177988052368, + 0.008619984611868858, + 0.1454421430826187, + 0.04358254373073578, + 0.006662610452622175, + 0.0068075405433773994, + 0.006995480041950941, + 0.02610652707517147, + 0.004593541845679283, + 0.005562165752053261, + 0.00930308923125267, + 0.015181466937065125, + 0.0025452389381825924, + 0.005298815201967955, + 0.004725499078631401, + 0.005668300203979015, + 0.003736478742212057, + 0.0016866844380274415, + 0.008086029440164566, + 0.006592464633285999, + 0.004956584423780441, + 0.006928985472768545, + 0.007436535321176052, + 0.00955044012516737, + 0.0043532815761864185, + 0.0064623733051121235, + 0.014208761975169182, + 0.010198189876973629, + 0.0031456556171178818, + 0.0037333995569497347, + 0.005028137471526861, + 0.004040922969579697, + 0.004721459001302719, + 0.004890452139079571, + 0.003229041351005435, + 0.0038650690112262964, + 0.002056267810985446, + 0.0030906759202480316, + 0.007735648658126593, + 0.002650260692462325, + 0.0012493041576817632, + 0.002041273983195424, + 0.002679546130821109, + 0.005597806069999933, + 0.0040087332017719746, + 0.005750427953898907, + 0.004018844570964575, + 0.004857274703681469, + 0.005717244930565357, + 0.004152534995228052, + 0.006304565351456404, + 0.0036354397889226675, + 0.014470060355961323, + 0.048977699130773544, + 0.01068995799869299, + 0.009561995975673199, + 0.00950669590383768, + 0.011431243270635605, + 0.022931693121790886, + 0.029707184061408043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "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 main subject an East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306214302778244 + ], + "opposite_logprob_tokens": [ + -3.168062210083008 + ], + "logprob_reward": -0.04306214302778244, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306214302778244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05407577008008957, + 0.04727330058813095, + 0.03964729979634285, + 0.513553261756897, + 0.031833868473768234, + 0.030972298234701157, + 0.025255249813199043, + 0.07912793010473251, + 0.05925467610359192, + 0.11900639533996582 + ], + "credit_map": [ + 0.00297604245133698, + 0.005679642781615257, + 0.0036271174903959036, + 0.016653239727020264, + 0.0052382261492311954, + 0.00408336753025651, + 0.012143825180828571, + 0.003674305509775877, + 0.004051298834383488, + 0.006284145172685385, + 0.00408795615658164, + 0.010638090781867504, + 0.005959595553576946, + 0.0023071153555065393, + 0.005872669164091349, + 0.008072431199252605, + 0.00192274060100317, + 0.010812538675963879, + 0.0028311219066381454, + 0.004033288452774286, + 0.0034307397436350584, + 0.0019945588428527117, + 0.004457410890609026, + 0.010164900682866573, + 0.029926754534244537, + 0.049558479338884354, + 0.021893562749028206, + 0.20989128947257996, + 0.02226167358458042, + 0.016470594331622124, + 0.06705280393362045, + 0.09649810194969177, + 0.0021783916745334864, + 0.005061264615505934, + 0.0041979714296758175, + 0.005387392360717058, + 0.0030982934404164553, + 0.0012318002991378307, + 0.005683073773980141, + 0.004995678085833788, + 0.002272221725434065, + 0.004433623980730772, + 0.003757740370929241, + 0.005581870675086975, + 0.0032473979517817497, + 0.0027452907525002956, + 0.004776191897690296, + 0.004157959949225187, + 0.0016442624619230628, + 0.0035644613672047853, + 0.004211832769215107, + 0.0039855181239545345, + 0.002115546725690365, + 0.002882427303120494, + 0.0024603696074336767, + 0.004390832036733627, + 0.0065061296336352825, + 0.006026384886354208, + 0.0223807692527771, + 0.0065409280359745026, + 0.004705587401986122, + 0.006340481340885162, + 0.00670101772993803, + 0.019926629960536957, + 0.004760147072374821, + 0.009657278656959534, + 0.004910114221274853, + 0.006847405340522528, + 0.005799204111099243, + 0.008301414549350739, + 0.014038700610399246, + 0.00494041433557868, + 0.010541032999753952, + 0.024149656295776367, + 0.008420969359576702, + 0.01255644392222166, + 0.011034952476620674, + 0.014223099686205387, + 0.018706178292632103, + 0.0193740576505661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.04306214302778244 + }, + { + "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 East Asian woman in her twenties with shoulder-length black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a charcoal gray turtleneck windbreaker?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a nighttime cityscape with motion-blurred streaks of electric blue, neon pink, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot from a slightly low-angle close-up with a backward tracking movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the woman's head held level while she walks forward with a steady, purposeful motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06629528850317001, + 0.13005463778972626, + 0.16655540466308594, + 0.2743825912475586, + 0.02017839439213276, + 0.02284197136759758, + 0.049535006284713745, + 0.04711497575044632, + 0.034042779356241226, + 0.18899892270565033 + ], + "credit_map": [ + 0.0032540957909077406, + 0.005210816860198975, + 0.00751307001337409, + 0.02017768658697605, + 0.006888974457979202, + 0.0039164102636277676, + 0.011946356855332851, + 0.007387877441942692, + 0.007404522970318794, + 0.014580674469470978, + 0.01390651986002922, + 0.030240289866924286, + 0.017663076519966125, + 0.006747062783688307, + 0.015618017874658108, + 0.023894470185041428, + 0.004353477619588375, + 0.054087746888399124, + 0.005279023200273514, + 0.007795017212629318, + 0.008191296830773354, + 0.0036012958735227585, + 0.0644034892320633, + 0.01884406805038452, + 0.015623376704752445, + 0.02617085725069046, + 0.010975614190101624, + 0.11185833811759949, + 0.011668872088193893, + 0.008992715738713741, + 0.035791970789432526, + 0.05330083891749382, + 0.0017469471786171198, + 0.0028876715805381536, + 0.0025858536828309298, + 0.0029660293366760015, + 0.0016256660455837846, + 0.001167057198472321, + 0.0035392697900533676, + 0.003659900277853012, + 0.0014600964495912194, + 0.005293217953294516, + 0.0022613529581576586, + 0.004193197935819626, + 0.0024294229224324226, + 0.001881964155472815, + 0.0026860074140131474, + 0.002636710414662957, + 0.004118798300623894, + 0.009892061352729797, + 0.006216972600668669, + 0.008856309577822685, + 0.002866547554731369, + 0.0057451422326266766, + 0.005582985933870077, + 0.006256192922592163, + 0.003647788893431425, + 0.004509565886110067, + 0.014184179715812206, + 0.005521310959011316, + 0.002852542093023658, + 0.003013844834640622, + 0.0038539557717740536, + 0.009531784802675247, + 0.002753069857135415, + 0.005198923870921135, + 0.0028115359600633383, + 0.004393344279378653, + 0.003658054629340768, + 0.005804511718451977, + 0.0061566527001559734, + 0.0032666849438101053, + 0.017500927671790123, + 0.03134540468454361, + 0.012937815859913826, + 0.031154559925198555, + 0.024452079087495804, + 0.0071090213023126125, + 0.05012759193778038, + 0.014371519908308983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.014195887371897697 + } + ], + "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/0000060/group_15/index_tensor([3365, 3365])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_tensor([3365, 3365])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..055026b752dfc9964c805f141d73b1e862cfba82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_tensor([3365, 3365])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Capturing the precise moment of takeoff with the breathtaking clarity of high-speed photography, the camera executes a smooth 360-degree orbit in ultra-slow motion around the majestic white crane. The scene begins with the crane, its perfect white feathers depicted in exquisite detail, startled from its perch atop a gnarled, ancient pine tree, its dark, cracked bark brimming with moisture. As its powerful wings, studded with black primary feathers, flap decisively downward, a wisp of glistening morning dew erupts from the pine needles. Each individual water droplet, suspended in the air, frozen in time like a tiny, multi-faceted gem, captures and refracts the warm golden light of the rising sun. Against the breathtaking backdrop of misty peaks, the orbiting camera captures the intricate texture of each feather, the powerful arch of its wings, and the brilliant prismatic light scattered through the water droplets, creating a serene and awe-inspiring atmosphere that freezes a fleeting moment of nature's power and exquisite beauty.", + "seed": 5977, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_3365_seed_5977_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": -0.49830890633165836, + "vlm_logprob": -0.49830890633165836, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.49830890633165836, + "vlm_yes_ratio": 0.6, + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8272445797920227 + ], + "opposite_logprob_tokens": [ + -0.5772445797920227 + ], + "logprob_reward": -0.8272445797920227, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8272445797920227, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08033781498670578, + 0.12508778274059296, + 0.056868426501750946, + 0.050852157175540924, + 0.0539800189435482, + 0.055194467306137085, + 0.1153271347284317, + 0.12403704226016998, + 0.06584060937166214, + 0.2724745571613312 + ], + "credit_map": [ + 0.007158199325203896, + 0.008306692354381084, + 0.010159600526094437, + 0.013746694661676884, + 0.008149220608174801, + 0.01382236834615469, + 0.010546431876718998, + 0.008448611944913864, + 0.009745164774358273, + 0.006579718552529812, + 0.0071236733347177505, + 0.008236792869865894, + 0.008857217617332935, + 0.040590979158878326, + 0.008990198373794556, + 0.03496403247117996, + 0.006998335011303425, + 0.005376200657337904, + 0.009388485923409462, + 0.0100996820256114, + 0.005410495679825544, + 0.005013330839574337, + 0.006785790901631117, + 0.007796108722686768, + 0.0062715583480894566, + 0.004812271799892187, + 0.007914094254374504, + 0.005215656943619251, + 0.00605663051828742, + 0.0053291902877390385, + 0.006363248452544212, + 0.008889502845704556, + 0.0035733513068407774, + 0.004099301062524319, + 0.0046187206171453, + 0.00636798981577158, + 0.005060113966464996, + 0.008057083934545517, + 0.005876387003809214, + 0.01632707379758358, + 0.0032306737266480923, + 0.0044383155182003975, + 0.007341848686337471, + 0.01016296911984682, + 0.00622808001935482, + 0.00827991683036089, + 0.010159979574382305, + 0.005352681968361139, + 0.0122005520388484, + 0.013389206491410732, + 0.016106408089399338, + 0.013760346919298172, + 0.016413873061537743, + 0.021114559844136238, + 0.01089441031217575, + 0.011447775177657604, + 0.007851166650652885, + 0.031670209020376205, + 0.010339498519897461, + 0.009934826754033566, + 0.008691072463989258, + 0.036005426198244095, + 0.009045513346791267, + 0.010499324649572372, + 0.0054810247384011745, + 0.011297171004116535, + 0.004060481674969196, + 0.008452973328530788, + 0.014342058449983597, + 0.008266718126833439, + 0.006447070278227329, + 0.007493110373616219, + 0.030518734827637672, + 0.03535904362797737, + 0.06443611532449722, + 0.03437504544854164, + 0.037667203694581985, + 0.03163296729326248, + 0.019108235836029053, + 0.019377201795578003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.8272445797920227 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.026365121826529503 + ], + "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": [ + 0.06272390484809875, + 0.18476742506027222, + 0.09337615966796875, + 0.19065618515014648, + 0.018728384748101234, + 0.113528773188591, + 0.10554434359073639, + 0.10699889063835144, + 0.024691130965948105, + 0.0989847257733345 + ], + "credit_map": [ + 0.005396571010351181, + 0.0056864297948777676, + 0.006332925986498594, + 0.00945961568504572, + 0.005675091408193111, + 0.006511214654892683, + 0.011590170674026012, + 0.012071889825165272, + 0.01636866107583046, + 0.010337070561945438, + 0.011875285767018795, + 0.015161865390837193, + 0.020156381651759148, + 0.04859980568289757, + 0.0165251512080431, + 0.04574321582913399, + 0.01631491631269455, + 0.009374778717756271, + 0.014506634324789047, + 0.010695323348045349, + 0.010943732224404812, + 0.008221732452511787, + 0.012426300905644894, + 0.010892733931541443, + 0.02631077729165554, + 0.017202438786625862, + 0.03584820032119751, + 0.01954597234725952, + 0.01830183155834675, + 0.01828213781118393, + 0.024216702207922935, + 0.030948126688599586, + 0.0016262710560113192, + 0.0017016124911606312, + 0.0024723948445171118, + 0.0028892983682453632, + 0.0021775385830551386, + 0.0016653976636007428, + 0.0028453280683606863, + 0.0033505430910736322, + 0.008084543980658054, + 0.03844446688890457, + 0.0094979926943779, + 0.011589931324124336, + 0.008130311965942383, + 0.007886930368840694, + 0.007719742599874735, + 0.022174853831529617, + 0.011862650513648987, + 0.010596995241940022, + 0.018852265551686287, + 0.013618088327348232, + 0.011968377977609634, + 0.01719949021935463, + 0.011293445713818073, + 0.010153031907975674, + 0.007812908850610256, + 0.031294841319322586, + 0.012170995585620403, + 0.009394079446792603, + 0.007592951413244009, + 0.024780338630080223, + 0.007584365084767342, + 0.006368412636220455, + 0.002714371308684349, + 0.003784327069297433, + 0.0020207890775054693, + 0.0025344707537442446, + 0.005029032938182354, + 0.0027261124923825264, + 0.002589769894257188, + 0.00329225673340261, + 0.010593481361865997, + 0.013759667053818703, + 0.018858416005969048, + 0.016147378832101822, + 0.011612145230174065, + 0.011270049028098583, + 0.008716112934052944, + 0.008027473464608192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "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 main subject a white crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.011080047115683556 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080047115683556, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080047115683556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04696662351489067, + 0.37963876128196716, + 0.030144406482577324, + 0.024764791131019592, + 0.05529113486409187, + 0.09195993095636368, + 0.08167222887277603, + 0.04474379122257233, + 0.04795157536864281, + 0.19686676561832428 + ], + "credit_map": [ + 0.0035852771252393723, + 0.004255971405655146, + 0.004169375170022249, + 0.006624176632612944, + 0.004343813750892878, + 0.0042994883842766285, + 0.008634889498353004, + 0.011053629219532013, + 0.019169388338923454, + 0.01482558436691761, + 0.01681767962872982, + 0.015982484444975853, + 0.024965155869722366, + 0.15029814839363098, + 0.022965913638472557, + 0.11461438983678818, + 0.0076480889692902565, + 0.0035315079148858786, + 0.00557375606149435, + 0.0021978761069476604, + 0.0024492223747074604, + 0.002392485039308667, + 0.0025793865788728, + 0.0037720822729170322, + 0.0040208082646131516, + 0.002570436103269458, + 0.004025012254714966, + 0.0026103234849870205, + 0.00242848158814013, + 0.0022313601803034544, + 0.002650626702234149, + 0.004227743484079838, + 0.003769687144085765, + 0.003586841281503439, + 0.0038388832472264767, + 0.0037659804802387953, + 0.007838994264602661, + 0.012501800432801247, + 0.004482252988964319, + 0.015506690368056297, + 0.006634828168898821, + 0.03275058791041374, + 0.006001123692840338, + 0.006345861125737429, + 0.006180755328387022, + 0.005990279372781515, + 0.007130623795092106, + 0.020925870165228844, + 0.009236964397132397, + 0.00852686446160078, + 0.01130069512873888, + 0.010230562649667263, + 0.01032214891165495, + 0.014607577584683895, + 0.008596417494118214, + 0.008851000107824802, + 0.004088213667273521, + 0.008121307939291, + 0.004447580315172672, + 0.0058518582955002785, + 0.005210882518440485, + 0.007389219477772713, + 0.005610013846307993, + 0.004024714231491089, + 0.005509148817509413, + 0.010055217891931534, + 0.003948654513806105, + 0.0038920750375837088, + 0.007721882313489914, + 0.005578604992479086, + 0.006147569045424461, + 0.005098419729620218, + 0.021624425426125526, + 0.024062585085630417, + 0.04216368868947029, + 0.026653600856661797, + 0.02832934446632862, + 0.026553066447377205, + 0.014794334769248962, + 0.012685724534094334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.011080047115683556 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.600504755973816 + ], + "opposite_logprob_tokens": [ + -0.22550472617149353 + ], + "logprob_reward": -1.600504755973816, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.600504755973816, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06615538895130157, + 0.3179774582386017, + 0.039295267313718796, + 0.0200121458619833, + 0.019695525988936424, + 0.06976515054702759, + 0.06552351266145706, + 0.15582814812660217, + 0.05017575994133949, + 0.195571631193161 + ], + "credit_map": [ + 0.00637465063482523, + 0.006238515954464674, + 0.007895119488239288, + 0.010780534707009792, + 0.0062100850045681, + 0.007376491092145443, + 0.01197749376296997, + 0.009302496910095215, + 0.014456747099757195, + 0.012846909463405609, + 0.014803479425609112, + 0.019930291920900345, + 0.02118583209812641, + 0.12474089115858078, + 0.018762631341814995, + 0.09125065803527832, + 0.00577189726755023, + 0.004207994788885117, + 0.005031990818679333, + 0.005533842835575342, + 0.003962715622037649, + 0.005608400795608759, + 0.003674149513244629, + 0.005504273809492588, + 0.0022133449092507362, + 0.002137264236807823, + 0.0027219823095947504, + 0.0033216204028576612, + 0.0015999984461814165, + 0.002576563972979784, + 0.002714501228183508, + 0.002726869424805045, + 0.0015627755783498287, + 0.0014043455012142658, + 0.0022217913065105677, + 0.0021153236739337444, + 0.002756214700639248, + 0.004462168086320162, + 0.0023403484374284744, + 0.0028325598686933517, + 0.004156470764428377, + 0.00986524112522602, + 0.010856795124709606, + 0.010006889700889587, + 0.005878114607185125, + 0.008878570050001144, + 0.0111481212079525, + 0.008974950760602951, + 0.007834436371922493, + 0.008449530228972435, + 0.006706908345222473, + 0.008373578079044819, + 0.007604590151458979, + 0.011374357156455517, + 0.007644847501069307, + 0.0075352671556174755, + 0.01586419716477394, + 0.04018924757838249, + 0.014054229483008385, + 0.01671055518090725, + 0.013950631022453308, + 0.028873765841126442, + 0.015561879612505436, + 0.010623627342283726, + 0.005795630160719156, + 0.006612150464206934, + 0.0037789333146065474, + 0.005574116948992014, + 0.010827518068253994, + 0.0058847940526902676, + 0.004625852685421705, + 0.007076763082295656, + 0.022401796653866768, + 0.013611187227070332, + 0.04672129452228546, + 0.02896454744040966, + 0.02278328873217106, + 0.02367660589516163, + 0.017965320497751236, + 0.019447600468993187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -1.600504755973816 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03462939336895943, + 0.5146613121032715, + 0.052016906440258026, + 0.03376098722219467, + 0.02510613389313221, + 0.038772016763687134, + 0.03751085326075554, + 0.11000091582536697, + 0.042846888303756714, + 0.11069460213184357 + ], + "credit_map": [ + 0.003121709218248725, + 0.0030623208731412888, + 0.00494383554905653, + 0.006388124544173479, + 0.002807486103847623, + 0.0037674815393984318, + 0.0062317028641700745, + 0.004306733142584562, + 0.02678886242210865, + 0.01967175118625164, + 0.02301759459078312, + 0.020762460306286812, + 0.03411641716957092, + 0.20438380539417267, + 0.029775375500321388, + 0.15614505112171173, + 0.008874493651092052, + 0.004499254748225212, + 0.007665510755032301, + 0.006385653279721737, + 0.0044329045340418816, + 0.005113303195685148, + 0.006303596775978804, + 0.008742189966142178, + 0.0056504919193685055, + 0.0026599913835525513, + 0.005807089153677225, + 0.0028555099852383137, + 0.0036533349193632603, + 0.0039919656701385975, + 0.0034343120642006397, + 0.005708294454962015, + 0.0022799137514084578, + 0.0018007337348535657, + 0.0024469054769724607, + 0.0024035917595028877, + 0.0031612024176865816, + 0.0035410502459853888, + 0.0026026777923107147, + 0.0068700578995049, + 0.0023483382537961006, + 0.011766946874558926, + 0.002788726706057787, + 0.0034680203534662724, + 0.003039811970666051, + 0.003930507227778435, + 0.0036777383647859097, + 0.00775193003937602, + 0.006603894289582968, + 0.00491162808611989, + 0.004812939558178186, + 0.004652153234928846, + 0.00447222450748086, + 0.00465382169932127, + 0.0035050229635089636, + 0.003899168921634555, + 0.005459959618747234, + 0.03430575132369995, + 0.011974128894507885, + 0.011018956080079079, + 0.007542488165199757, + 0.02648906409740448, + 0.007110775914043188, + 0.006099783815443516, + 0.0038283520843833685, + 0.007988259196281433, + 0.003461485495790839, + 0.0043871356174349785, + 0.00824839249253273, + 0.005277970805764198, + 0.0047814566642045975, + 0.004873835947364569, + 0.011588254943490028, + 0.012724623084068298, + 0.025196705013513565, + 0.0183024350553751, + 0.009847791865468025, + 0.008764898404479027, + 0.01018605101853609, + 0.014083846472203732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.026350026950240135 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Capturing the precise moment of takeoff with the breathtaking clarity of high-speed photography, the camera executes a smooth 360-degree orbit in ultra-slow motion around the majestic white crane. The scene begins with the crane, its perfect white feathers depicted in exquisite detail, startled from its perch atop a gnarled, ancient pine tree, its dark, cracked bark brimming with moisture. As its powerful wings, studded with black primary feathers, flap decisively downward, a wisp of glistening morning dew erupts from the pine needles. Each individual water droplet, suspended in the air, frozen in time like a tiny, multi-faceted gem, captures and refracts the warm golden light of the rising sun. Against the breathtaking backdrop of misty peaks, the orbiting camera captures the intricate texture of each feather, the powerful arch of its wings, and the brilliant prismatic light scattered through the water droplets, creating a serene and awe-inspiring atmosphere that freezes a fleeting moment of nature's power and exquisite beauty.", + "seed": 5978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_3365_seed_5978_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": -0.47379800006747247, + "vlm_logprob": -0.47379800006747247, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.47379800006747247, + "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 main subject a white crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.47734397649765015 + ], + "opposite_logprob_tokens": [ + -0.9773439764976501 + ], + "logprob_reward": -0.47734397649765015, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47734397649765015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1231575459241867, + 0.048636503517627716, + 0.03262612223625183, + 0.07242913544178009, + 0.21276096999645233, + 0.06451444327831268, + 0.08009928464889526, + 0.04049203544855118, + 0.14774224162101746, + 0.17754165828227997 + ], + "credit_map": [ + 0.014100853353738785, + 0.0191374309360981, + 0.011915817856788635, + 0.016827162355184555, + 0.01311139203608036, + 0.011505277827382088, + 0.01959858648478985, + 0.016961034387350082, + 0.004267687909305096, + 0.0032713545951992273, + 0.006034241523593664, + 0.0063368165865540504, + 0.008661643601953983, + 0.005720000248402357, + 0.007089036982506514, + 0.00725572369992733, + 0.002266254276037216, + 0.002591808093711734, + 0.0036651436239480972, + 0.005693352781236172, + 0.003818992292508483, + 0.0026442536618560553, + 0.006225052755326033, + 0.005721268709748983, + 0.003489691298455, + 0.003874234389513731, + 0.00429549952968955, + 0.004209810867905617, + 0.0027576747816056013, + 0.01622220315039158, + 0.00998047087341547, + 0.027599556371569633, + 0.025533750653266907, + 0.03353847935795784, + 0.024793967604637146, + 0.03666241466999054, + 0.028386589139699936, + 0.01693510264158249, + 0.01984531432390213, + 0.027065366506576538, + 0.005087795667350292, + 0.007025451399385929, + 0.0069193365052342415, + 0.009063122794032097, + 0.010367139242589474, + 0.00613574031740427, + 0.009465469978749752, + 0.01045039389282465, + 0.00841030664741993, + 0.009915731847286224, + 0.00808279775083065, + 0.012790670618414879, + 0.010095431469380856, + 0.006625934969633818, + 0.011686841957271099, + 0.01249157264828682, + 0.0058408621698617935, + 0.005505229812115431, + 0.004035120364278555, + 0.0036744505632668734, + 0.006371107883751392, + 0.005264041014015675, + 0.003832233836874366, + 0.005968993995338678, + 0.011005764827132225, + 0.01090187393128872, + 0.026237988844513893, + 0.011955928057432175, + 0.02153170481324196, + 0.04117884859442711, + 0.015515133738517761, + 0.009415008127689362, + 0.022051315754652023, + 0.01870732381939888, + 0.02667732909321785, + 0.02839214913547039, + 0.019527196884155273, + 0.017570247873663902, + 0.026841681450605392, + 0.017774416133761406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.47734397649765015 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.029818223789334297 + ], + "opposite_logprob_tokens": [ + -3.529818296432495 + ], + "logprob_reward": -0.029818223789334297, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029818223789334297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2606200575828552, + 0.04878566041588783, + 0.07001333683729172, + 0.05425083264708519, + 0.3262287378311157, + 0.05720416083931923, + 0.024717792868614197, + 0.02313680201768875, + 0.06341738998889923, + 0.07162530720233917 + ], + "credit_map": [ + 0.032038502395153046, + 0.04934249445796013, + 0.03316900134086609, + 0.02840178832411766, + 0.0263492651283741, + 0.02033640258014202, + 0.030717860907316208, + 0.04026472568511963, + 0.004717777017503977, + 0.004646620247513056, + 0.007616586983203888, + 0.008083981461822987, + 0.004534157924354076, + 0.0036369748413562775, + 0.007481391541659832, + 0.008068174123764038, + 0.006454013753682375, + 0.006095643155276775, + 0.01043104287236929, + 0.010731947608292103, + 0.00949084758758545, + 0.005829673260450363, + 0.011323517188429832, + 0.009656657464802265, + 0.0055291615426540375, + 0.008896386250853539, + 0.006344211753457785, + 0.0071084098890423775, + 0.005185925867408514, + 0.004092637449502945, + 0.009815174154937267, + 0.007278922945261002, + 0.04096429422497749, + 0.04398581385612488, + 0.04026559740304947, + 0.05188879743218422, + 0.042486246675252914, + 0.026858393102884293, + 0.032460007816553116, + 0.04731958732008934, + 0.00479678250849247, + 0.007905067875981331, + 0.006140486337244511, + 0.008414456620812416, + 0.0084693543612957, + 0.005234411917626858, + 0.008901764638721943, + 0.007341837510466576, + 0.002265371847897768, + 0.0029201367869973183, + 0.002364370273426175, + 0.0033652957063168287, + 0.003317935159429908, + 0.0024348939768970013, + 0.004121161065995693, + 0.0039286259561777115, + 0.003057132475078106, + 0.0027671006973832846, + 0.0030625739600509405, + 0.0029472731985151768, + 0.0029731837566941977, + 0.0025705278385430574, + 0.002486692974343896, + 0.003272317349910736, + 0.005293954163789749, + 0.006198944058269262, + 0.0098270820453763, + 0.004912905860692263, + 0.00823983270674944, + 0.013663147576153278, + 0.009525610134005547, + 0.005755915306508541, + 0.008995302952826023, + 0.009050069376826286, + 0.01164073683321476, + 0.008618419989943504, + 0.00563051225617528, + 0.007181383669376373, + 0.011785615235567093, + 0.008723261766135693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.029818223789334297 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.023295411840081215 + ], + "opposite_logprob_tokens": [ + -3.7732954025268555 + ], + "logprob_reward": -0.023295411840081215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295411840081215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1555280089378357, + 0.0783868059515953, + 0.06785069406032562, + 0.0620308518409729, + 0.1502889096736908, + 0.054045528173446655, + 0.05229892581701279, + 0.042569708079099655, + 0.10700716078281403, + 0.22999338805675507 + ], + "credit_map": [ + 0.01576812192797661, + 0.015461149625480175, + 0.012779992073774338, + 0.018052108585834503, + 0.028858620673418045, + 0.017080744728446007, + 0.023115307092666626, + 0.024411974474787712, + 0.005650491453707218, + 0.006133066490292549, + 0.007123176008462906, + 0.015060474164783955, + 0.008819909766316414, + 0.006748392712324858, + 0.015726367011666298, + 0.013124938122928143, + 0.0048769256100058556, + 0.010179626755416393, + 0.009422242641448975, + 0.011658123694360256, + 0.009756242856383324, + 0.005439870059490204, + 0.007558007724583149, + 0.008959664963185787, + 0.005011457949876785, + 0.005688697099685669, + 0.0055190399289131165, + 0.005508490838110447, + 0.003937819041311741, + 0.011282183229923248, + 0.008745196275413036, + 0.016337968409061432, + 0.017603660002350807, + 0.02083687111735344, + 0.018739772960543633, + 0.021995771676301956, + 0.01934419944882393, + 0.012062904424965382, + 0.021377084776759148, + 0.018328659236431122, + 0.0038710213266313076, + 0.006969582289457321, + 0.006170632783323526, + 0.006600040476769209, + 0.006799187511205673, + 0.004504954908043146, + 0.00837015826255083, + 0.010759955272078514, + 0.0033591848332434893, + 0.005165452137589455, + 0.00655405642464757, + 0.00896407663822174, + 0.00594515772536397, + 0.004099379293620586, + 0.007598597556352615, + 0.010613025166094303, + 0.005153122823685408, + 0.005365604069083929, + 0.00578986806795001, + 0.003625094424933195, + 0.007282676640897989, + 0.0042001958936452866, + 0.00507388124242425, + 0.006079265382140875, + 0.010160040110349655, + 0.010563432238996029, + 0.018005531281232834, + 0.00833918061107397, + 0.012261681258678436, + 0.026759665459394455, + 0.013145839795470238, + 0.007771792821586132, + 0.030614910647273064, + 0.027503540739417076, + 0.03417166694998741, + 0.03159908577799797, + 0.02915956825017929, + 0.019094733521342278, + 0.03900190815329552, + 0.018847988918423653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.023295411840081215 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048250675201416 + ], + "opposite_logprob_tokens": [ + -0.17982502281665802 + ], + "logprob_reward": -1.8048250675201416, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048250675201416, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1391209065914154, + 0.07295212149620056, + 0.05666728690266609, + 0.07890989631414413, + 0.09868481755256653, + 0.07861407101154327, + 0.09868059307336807, + 0.033915672451257706, + 0.13666795194149017, + 0.20578664541244507 + ], + "credit_map": [ + 0.014927777461707592, + 0.011756918393075466, + 0.013948339968919754, + 0.023395754396915436, + 0.012020908296108246, + 0.011336103081703186, + 0.023262890055775642, + 0.028472209349274635, + 0.006445170845836401, + 0.006680727936327457, + 0.006482382770627737, + 0.013103259727358818, + 0.008951284922659397, + 0.006736586336046457, + 0.012612530030310154, + 0.011940181255340576, + 0.005833916366100311, + 0.008046181872487068, + 0.0071182576939463615, + 0.01145028043538332, + 0.006438622251152992, + 0.0036799197550863028, + 0.007036883383989334, + 0.007063225377351046, + 0.0064622038044035435, + 0.006201886106282473, + 0.006681504659354687, + 0.005032539367675781, + 0.004981544334441423, + 0.013836989179253578, + 0.012003792449831963, + 0.023709433153271675, + 0.011935208924114704, + 0.01064921822398901, + 0.0101200882345438, + 0.017676951363682747, + 0.012047551572322845, + 0.007809323258697987, + 0.013315066695213318, + 0.015131407417356968, + 0.007319573312997818, + 0.005872371140867472, + 0.010483082383871078, + 0.011389156803488731, + 0.012784196995198727, + 0.0059144156984984875, + 0.015932604670524597, + 0.008918668143451214, + 0.012201152741909027, + 0.012555219233036041, + 0.010743314400315285, + 0.014806807972490788, + 0.01247079111635685, + 0.009478711523115635, + 0.010716688819229603, + 0.01570790261030197, + 0.004388036206364632, + 0.005107898730784655, + 0.004379910882562399, + 0.003739582374691963, + 0.004305517766624689, + 0.0032786254305392504, + 0.0036254532169550657, + 0.005090648308396339, + 0.016674574464559555, + 0.013809568248689175, + 0.016424080356955528, + 0.01446497905999422, + 0.01439975667744875, + 0.02256781794130802, + 0.021873364225029945, + 0.016453806310892105, + 0.016970735043287277, + 0.023638198152184486, + 0.024409683421254158, + 0.023264486342668533, + 0.05131957307457924, + 0.026178253814578056, + 0.021184151992201805, + 0.01882157102227211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -1.8048250675201416 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16095739603042603, + 0.04227204993367195, + 0.059111423790454865, + 0.1321536898612976, + 0.08939994871616364, + 0.04367784038186073, + 0.03285779803991318, + 0.03764137253165245, + 0.21598833799362183, + 0.18594014644622803 + ], + "credit_map": [ + 0.0153351416811347, + 0.024157434701919556, + 0.015046787448227406, + 0.017939792945981026, + 0.030224015936255455, + 0.01695280149579048, + 0.018675651401281357, + 0.022625768557190895, + 0.002934612100943923, + 0.0038601390551775694, + 0.00578111968934536, + 0.005634626839309931, + 0.005992885213345289, + 0.0037238062359392643, + 0.007266401778906584, + 0.007078460417687893, + 0.005917892791330814, + 0.005987551994621754, + 0.008627722039818764, + 0.006244841031730175, + 0.008137152530252934, + 0.004861437249928713, + 0.009892293252050877, + 0.009442533366382122, + 0.004562925547361374, + 0.003751736367121339, + 0.00439640274271369, + 0.005219748243689537, + 0.0041848537512123585, + 0.03541756421327591, + 0.017517214640975, + 0.0571032352745533, + 0.009444834664463997, + 0.011837318539619446, + 0.010288368910551071, + 0.012277010828256607, + 0.013730017468333244, + 0.007436203770339489, + 0.012456504628062248, + 0.011929689906537533, + 0.003212924348190427, + 0.00522932643070817, + 0.004662682302296162, + 0.00672593479976058, + 0.007330520078539848, + 0.003396357409656048, + 0.006532975938171148, + 0.006587118376046419, + 0.0020456493366509676, + 0.002672898815944791, + 0.004220993258059025, + 0.005234906449913979, + 0.0032839516643434763, + 0.0028686155565083027, + 0.006537515204399824, + 0.0059932684525847435, + 0.005304075311869383, + 0.005162304732948542, + 0.0058862194418907166, + 0.003568003885447979, + 0.00498398020863533, + 0.004337073769420385, + 0.003220371436327696, + 0.005179343279451132, + 0.01747726835310459, + 0.02161034196615219, + 0.03995350003242493, + 0.015776697546243668, + 0.02973151206970215, + 0.06176215410232544, + 0.01680675335228443, + 0.012870111502707005, + 0.021062245592474937, + 0.024927474558353424, + 0.023604054003953934, + 0.03292715176939964, + 0.028305795043706894, + 0.01949230581521988, + 0.023174891248345375, + 0.012446223758161068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.03370732069015503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Capturing the precise moment of takeoff with the breathtaking clarity of high-speed photography, the camera executes a smooth 360-degree orbit in ultra-slow motion around the majestic white crane. The scene begins with the crane, its perfect white feathers depicted in exquisite detail, startled from its perch atop a gnarled, ancient pine tree, its dark, cracked bark brimming with moisture. As its powerful wings, studded with black primary feathers, flap decisively downward, a wisp of glistening morning dew erupts from the pine needles. Each individual water droplet, suspended in the air, frozen in time like a tiny, multi-faceted gem, captures and refracts the warm golden light of the rising sun. Against the breathtaking backdrop of misty peaks, the orbiting camera captures the intricate texture of each feather, the powerful arch of its wings, and the brilliant prismatic light scattered through the water droplets, creating a serene and awe-inspiring atmosphere that freezes a fleeting moment of nature's power and exquisite beauty.", + "seed": 5979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_3365_seed_5979_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": -0.5903597220778465, + "vlm_logprob": -0.5903597220778465, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.5903597220778465, + "vlm_yes_ratio": 0.6, + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9773851037025452 + ], + "opposite_logprob_tokens": [ + -0.47738510370254517 + ], + "logprob_reward": -0.9773851037025452, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9773851037025452, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10555700212717056, + 0.0572112500667572, + 0.08119969815015793, + 0.08240704238414764, + 0.0958794429898262, + 0.03536870330572128, + 0.05232493206858635, + 0.3680126667022705, + 0.05819719657301903, + 0.06384206563234329 + ], + "credit_map": [ + 0.0193867776542902, + 0.025440383702516556, + 0.010034956969320774, + 0.009961381554603577, + 0.012665851972997189, + 0.010441569611430168, + 0.008360196836292744, + 0.009265889413654804, + 0.008309323340654373, + 0.007223994936794043, + 0.00461022462695837, + 0.008040777407586575, + 0.007594429887831211, + 0.009554635733366013, + 0.003989450167864561, + 0.007888412103056908, + 0.013329114764928818, + 0.011691104620695114, + 0.006057718303054571, + 0.006956118158996105, + 0.013749946840107441, + 0.012002721428871155, + 0.00881236046552658, + 0.008600616827607155, + 0.013291370123624802, + 0.01504873763769865, + 0.005828173831105232, + 0.008433574810624123, + 0.014251457527279854, + 0.010215495713055134, + 0.007837517186999321, + 0.007500709965825081, + 0.014407621696591377, + 0.013399014249444008, + 0.010260995477437973, + 0.019048180431127548, + 0.013065418228507042, + 0.009083454497158527, + 0.0066747041419148445, + 0.009940054267644882, + 0.0049321637488901615, + 0.0036247193347662687, + 0.004465139005333185, + 0.005137366242706776, + 0.007552851922810078, + 0.003820674493908882, + 0.0029601347632706165, + 0.0028756523970514536, + 0.006247301120311022, + 0.004329835530370474, + 0.010159237310290337, + 0.004475697875022888, + 0.013431739062070847, + 0.0041229999624192715, + 0.007323724217712879, + 0.0022343972232192755, + 0.03951939195394516, + 0.023426763713359833, + 0.07975514233112335, + 0.034742116928100586, + 0.12117340415716171, + 0.03258206322789192, + 0.023672420531511307, + 0.013141353614628315, + 0.006377904675900936, + 0.004751077853143215, + 0.009403562173247337, + 0.007378008682280779, + 0.008109905757009983, + 0.007722760085016489, + 0.009229464456439018, + 0.005224511958658695, + 0.004695548675954342, + 0.003114157123491168, + 0.020461559295654297, + 0.017131319269537926, + 0.0030824660789221525, + 0.0039289044216275215, + 0.008001384325325489, + 0.003426726907491684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.9773851037025452 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.020610567182302475 + ], + "opposite_logprob_tokens": [ + -3.8956105709075928 + ], + "logprob_reward": -0.020610567182302475, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610567182302475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11000440269708633, + 0.04938844218850136, + 0.19086825847625732, + 0.05518374592065811, + 0.16040536761283875, + 0.04717310890555382, + 0.09436193108558655, + 0.21450650691986084, + 0.05469539016485214, + 0.023412887006998062 + ], + "credit_map": [ + 0.025759320706129074, + 0.03034573420882225, + 0.011935187503695488, + 0.008548426441848278, + 0.009974471293389797, + 0.007736737374216318, + 0.00682271970435977, + 0.00888178963214159, + 0.006893982645124197, + 0.007166625931859016, + 0.00460826326161623, + 0.007122304290533066, + 0.006302503868937492, + 0.005488474387675524, + 0.0046882363967597485, + 0.0071180458180606365, + 0.030571766197681427, + 0.020045790821313858, + 0.016021769493818283, + 0.016626037657260895, + 0.02983078360557556, + 0.03055943176150322, + 0.02248322032392025, + 0.024729430675506592, + 0.008261604234576225, + 0.008141374215483665, + 0.004425001330673695, + 0.005551676265895367, + 0.007506216876208782, + 0.008815201930701733, + 0.005561217200011015, + 0.006921445485204458, + 0.02470262721180916, + 0.029692040756344795, + 0.01739141345024109, + 0.030869990587234497, + 0.020917858928442, + 0.011162443086504936, + 0.011064691469073296, + 0.01460428535938263, + 0.00710564199835062, + 0.006926822476089001, + 0.005805891007184982, + 0.0062040467746555805, + 0.007948718033730984, + 0.004140640143305063, + 0.0039204987697303295, + 0.005120846442878246, + 0.023041127249598503, + 0.00941326841711998, + 0.0061067803762853146, + 0.003654356114566326, + 0.033427488058805466, + 0.008523172698915005, + 0.0052568051032722, + 0.004938922822475433, + 0.022461768239736557, + 0.01855292357504368, + 0.0288043562322855, + 0.01959477923810482, + 0.04585041105747223, + 0.029851317405700684, + 0.018334202468395233, + 0.031056737527251244, + 0.004438414238393307, + 0.002514938823878765, + 0.01692013442516327, + 0.014525271020829678, + 0.003299347823485732, + 0.004000921733677387, + 0.006319537293165922, + 0.0026768262032419443, + 0.001915237633511424, + 0.0019755945540964603, + 0.004578702617436647, + 0.004523841664195061, + 0.002125783124938607, + 0.0023468579165637493, + 0.003433810081332922, + 0.002513057552278042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.020610567182302475 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.009791914373636246 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791914373636246, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791914373636246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2529636025428772, + 0.04954787716269493, + 0.12492042779922485, + 0.031394053250551224, + 0.14325855672359467, + 0.0485575906932354, + 0.05594924837350845, + 0.18592491745948792, + 0.045141227543354034, + 0.06234255060553551 + ], + "credit_map": [ + 0.06622336059808731, + 0.07834326475858688, + 0.020808828994631767, + 0.01745670475065708, + 0.023029152303934097, + 0.016686635091900826, + 0.011707567609846592, + 0.01870810054242611, + 0.006322621367871761, + 0.005994592327624559, + 0.005662995856255293, + 0.008994495496153831, + 0.006016288883984089, + 0.00567292096093297, + 0.005300414748489857, + 0.005583550781011581, + 0.019617129117250443, + 0.025380872189998627, + 0.011390435509383678, + 0.01804473251104355, + 0.017942728474736214, + 0.012535841204226017, + 0.010411054827272892, + 0.009597641415894032, + 0.003718384774401784, + 0.005561182741075754, + 0.002396446419879794, + 0.002653328236192465, + 0.005667387042194605, + 0.004940895363688469, + 0.002703895093873143, + 0.0037525342777371407, + 0.02292373590171337, + 0.03326570987701416, + 0.012967647053301334, + 0.025710292160511017, + 0.01994461566209793, + 0.00858041550964117, + 0.009805934503674507, + 0.010060210712254047, + 0.0064357309602200985, + 0.010106723755598068, + 0.003935119602829218, + 0.005537087097764015, + 0.006996818818151951, + 0.006058407016098499, + 0.005068251397460699, + 0.004419451579451561, + 0.0068351756781339645, + 0.010264933109283447, + 0.004294732119888067, + 0.006148352287709713, + 0.007754855789244175, + 0.005476115271449089, + 0.011404475197196007, + 0.003770607989281416, + 0.018949855118989944, + 0.009862043894827366, + 0.041149627417325974, + 0.020140398293733597, + 0.05963417887687683, + 0.01634768396615982, + 0.012009654194116592, + 0.007831468246877193, + 0.004115581046789885, + 0.0033827335573732853, + 0.007991134189069271, + 0.00688112573698163, + 0.005849223118275404, + 0.005802781321108341, + 0.007194183301180601, + 0.0039244648069143295, + 0.006189906504005194, + 0.005323162768036127, + 0.015685802325606346, + 0.01215439010411501, + 0.0051098838448524475, + 0.004836502950638533, + 0.007919536903500557, + 0.005123363807797432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.009791914373636246 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103000164031982 + ], + "opposite_logprob_tokens": [ + -0.16030006110668182 + ], + "logprob_reward": -1.9103000164031982, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103000164031982, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11461349576711655, + 0.08681557327508926, + 0.057948704808950424, + 0.25162363052368164, + 0.04158306121826172, + 0.05159301683306694, + 0.04783244803547859, + 0.09262912720441818, + 0.06869827955961227, + 0.18666264414787292 + ], + "credit_map": [ + 0.0231680478900671, + 0.028531521558761597, + 0.010622229427099228, + 0.0091191865503788, + 0.01366609800606966, + 0.013316301628947258, + 0.006548129487782717, + 0.009641987271606922, + 0.009878224693238735, + 0.01146264374256134, + 0.008912111632525921, + 0.011739782057702541, + 0.01142655499279499, + 0.011600836180150509, + 0.009314649738371372, + 0.012480762787163258, + 0.008874035440385342, + 0.008744641207158566, + 0.005114064086228609, + 0.00751875014975667, + 0.0069892518222332, + 0.005956039298325777, + 0.005740936845541, + 0.00901099108159542, + 0.039040934294462204, + 0.03389643877744675, + 0.018224680796265602, + 0.031155386939644814, + 0.04172569885849953, + 0.03857870399951935, + 0.024279886856675148, + 0.024721907451748848, + 0.006844220217317343, + 0.01284072082489729, + 0.004380804020911455, + 0.00414590910077095, + 0.004188441205769777, + 0.0024862892460078, + 0.00268634338863194, + 0.004010334145277739, + 0.006986386142671108, + 0.011738123372197151, + 0.004423238802701235, + 0.006935144308954477, + 0.006229252088814974, + 0.005604192614555359, + 0.00462089478969574, + 0.00505578238517046, + 0.006153248716145754, + 0.0033713600132614374, + 0.011131917126476765, + 0.004206948447972536, + 0.012530248612165451, + 0.0038056259509176016, + 0.004569428972899914, + 0.002063667867332697, + 0.010436558164656162, + 0.006528666242957115, + 0.019999144598841667, + 0.010581087321043015, + 0.02490907721221447, + 0.008442377671599388, + 0.006634300108999014, + 0.005097921937704086, + 0.005506834480911493, + 0.004237317480146885, + 0.020226243883371353, + 0.01837955228984356, + 0.003956985659897327, + 0.005087010562419891, + 0.00717394333332777, + 0.004130394198000431, + 0.017951391637325287, + 0.01595245487987995, + 0.032807473093271255, + 0.03669635206460953, + 0.02311982586979866, + 0.018541650846600533, + 0.0257217139005661, + 0.015871796756982803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -1.9103000164031982 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.03371100872755051 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.03371100872755051, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371100872755051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19005565345287323, + 0.07163874804973602, + 0.10392861068248749, + 0.20251019299030304, + 0.05199331417679787, + 0.06647422164678574, + 0.07776952534914017, + 0.056452441960573196, + 0.03292218968272209, + 0.14625512063503265 + ], + "credit_map": [ + 0.04015630856156349, + 0.04630647227168083, + 0.014947673305869102, + 0.012803713791072369, + 0.024596402421593666, + 0.026731321588158607, + 0.010558812879025936, + 0.013954957015812397, + 0.008249599486589432, + 0.009609528817236423, + 0.007847451604902744, + 0.009499093517661095, + 0.009883096441626549, + 0.010104666464030743, + 0.0076193008571863174, + 0.008826009929180145, + 0.015852918848395348, + 0.014056550338864326, + 0.010643727146089077, + 0.010779477655887604, + 0.01363637950271368, + 0.01334582082927227, + 0.008698910474777222, + 0.016914814710617065, + 0.03095187060534954, + 0.02652093768119812, + 0.014057167805731297, + 0.022768845781683922, + 0.03712347149848938, + 0.03510299324989319, + 0.017155474051833153, + 0.018829425796866417, + 0.008153926581144333, + 0.012926837429404259, + 0.005349103361368179, + 0.006256915628910065, + 0.006239297799766064, + 0.0035341400653123856, + 0.0037818208802491426, + 0.005751274526119232, + 0.008360306732356548, + 0.005618825554847717, + 0.006970117799937725, + 0.006410328671336174, + 0.017758114263415337, + 0.007849900983273983, + 0.009311538189649582, + 0.004195091314613819, + 0.014555973932147026, + 0.00452874880284071, + 0.013024774380028248, + 0.0052339499816298485, + 0.028867337852716446, + 0.004895736929029226, + 0.0037745372392237186, + 0.0028884699568152428, + 0.00742596248164773, + 0.006851060315966606, + 0.006281625479459763, + 0.006993670482188463, + 0.00668865954503417, + 0.007558076176792383, + 0.007688646204769611, + 0.0069647375494241714, + 0.0036021883133798838, + 0.0022238309029489756, + 0.007864819839596748, + 0.006935253739356995, + 0.0028643289115279913, + 0.0023475377820432186, + 0.00508371414616704, + 0.0020005186088383198, + 0.014768811874091625, + 0.011504320427775383, + 0.034514229744672775, + 0.028084933757781982, + 0.01671101339161396, + 0.01351096760481596, + 0.016214773058891296, + 0.01094607263803482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.03371100872755051 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Capturing the precise moment of takeoff with the breathtaking clarity of high-speed photography, the camera executes a smooth 360-degree orbit in ultra-slow motion around the majestic white crane. The scene begins with the crane, its perfect white feathers depicted in exquisite detail, startled from its perch atop a gnarled, ancient pine tree, its dark, cracked bark brimming with moisture. As its powerful wings, studded with black primary feathers, flap decisively downward, a wisp of glistening morning dew erupts from the pine needles. Each individual water droplet, suspended in the air, frozen in time like a tiny, multi-faceted gem, captures and refracts the warm golden light of the rising sun. Against the breathtaking backdrop of misty peaks, the orbiting camera captures the intricate texture of each feather, the powerful arch of its wings, and the brilliant prismatic light scattered through the water droplets, creating a serene and awe-inspiring atmosphere that freezes a fleeting moment of nature's power and exquisite beauty.", + "seed": 5980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_15/index_3365_seed_5980_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": -0.6319913763552905, + "vlm_logprob": -0.6319913763552905, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.6319913763552905, + "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 main subject a white crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.39370450377464294 + ], + "opposite_logprob_tokens": [ + -1.1437045335769653 + ], + "logprob_reward": -0.39370450377464294, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39370450377464294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17204676568508148, + 0.10902514308691025, + 0.03843310847878456, + 0.05244661122560501, + 0.04345466196537018, + 0.06345880031585693, + 0.03251684457063675, + 0.03076694905757904, + 0.3721660375595093, + 0.0856851115822792 + ], + "credit_map": [ + 0.027355458587408066, + 0.02179580181837082, + 0.033735912293195724, + 0.01847679913043976, + 0.02696465514600277, + 0.01540627796202898, + 0.015047269873321056, + 0.01326458714902401, + 0.009836604818701744, + 0.013160213828086853, + 0.027083443477749825, + 0.00913409423083067, + 0.016839871183037758, + 0.01316208764910698, + 0.009085455909371376, + 0.01072336733341217, + 0.0037268237210810184, + 0.008165997453033924, + 0.006406250409781933, + 0.004371990449726582, + 0.005502788349986076, + 0.004659289959818125, + 0.003077938687056303, + 0.0025220289826393127, + 0.004829510115087032, + 0.008082395419478416, + 0.006926416885107756, + 0.004978945013135672, + 0.006495985668152571, + 0.012056657113134861, + 0.00518243620172143, + 0.003894262248650193, + 0.005118578672409058, + 0.005554411094635725, + 0.005031378474086523, + 0.004400153178721666, + 0.008435703814029694, + 0.00671097869053483, + 0.004581338260322809, + 0.0036221230402588844, + 0.007143912371248007, + 0.007872418500483036, + 0.008328784257173538, + 0.005883566569536924, + 0.010270481929183006, + 0.006778509821742773, + 0.010413297452032566, + 0.0067678228951990604, + 0.0029026141855865717, + 0.005983904004096985, + 0.004016842693090439, + 0.005112052895128727, + 0.005413406994193792, + 0.002187544945627451, + 0.003491748357191682, + 0.003408733755350113, + 0.0028865844942629337, + 0.006007552146911621, + 0.003511752001941204, + 0.002982131438329816, + 0.006012719124555588, + 0.003934039734303951, + 0.002401767997071147, + 0.003030401421710849, + 0.05103302001953125, + 0.08878541737794876, + 0.046303559094667435, + 0.04586569964885712, + 0.042621415108442307, + 0.03737463057041168, + 0.0281058419495821, + 0.03207650035619736, + 0.010615775361657143, + 0.013940898701548576, + 0.008375479839742184, + 0.01738874986767769, + 0.009032807312905788, + 0.008165881969034672, + 0.009312373585999012, + 0.00885313656181097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.39370450377464294 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.020603211596608162 + ], + "opposite_logprob_tokens": [ + -3.8956031799316406 + ], + "logprob_reward": -0.020603211596608162, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603211596608162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3843940198421478, + 0.13641537725925446, + 0.02837972901761532, + 0.1020912304520607, + 0.04958255589008331, + 0.06013735383749008, + 0.03410329669713974, + 0.05177650228142738, + 0.053160786628723145, + 0.09995909035205841 + ], + "credit_map": [ + 0.06520936638116837, + 0.05738957226276398, + 0.06779596209526062, + 0.029566818848252296, + 0.06615537405014038, + 0.03606952354311943, + 0.031592000275850296, + 0.030615437775850296, + 0.012056759558618069, + 0.016500622034072876, + 0.031955014914274216, + 0.016602275893092155, + 0.017215624451637268, + 0.013555902056396008, + 0.015242474153637886, + 0.013286720030009747, + 0.003907144535332918, + 0.0036793581675738096, + 0.0060372608713805676, + 0.0025731169153004885, + 0.003977314103394747, + 0.002736867405474186, + 0.0026785850059241056, + 0.0027900852728635073, + 0.01068247202783823, + 0.01777520403265953, + 0.016415340825915337, + 0.008895599283277988, + 0.01975408010184765, + 0.016353197395801544, + 0.006976834032684565, + 0.005238513927906752, + 0.008259668946266174, + 0.005969599355012178, + 0.00530308298766613, + 0.005222444422543049, + 0.005963452626019716, + 0.007694800850003958, + 0.006024043075740337, + 0.005145464092493057, + 0.010397781617939472, + 0.009650517255067825, + 0.007890475913882256, + 0.005045230034738779, + 0.008938522078096867, + 0.005270304623991251, + 0.008007396943867207, + 0.004937130957841873, + 0.003193905809894204, + 0.0045080604031682014, + 0.004221588373184204, + 0.0054841977544128895, + 0.005033448338508606, + 0.002284625079482794, + 0.005396486725658178, + 0.00398098723962903, + 0.005553610157221556, + 0.006153928115963936, + 0.006414232309907675, + 0.007107592187821865, + 0.010884839110076427, + 0.006964321248233318, + 0.004129806067794561, + 0.004568174481391907, + 0.006339100189507008, + 0.007937091402709484, + 0.006900605745613575, + 0.006453700363636017, + 0.00598579365760088, + 0.009014621376991272, + 0.004508474376052618, + 0.006021400913596153, + 0.016935313120484352, + 0.01616864837706089, + 0.011446597054600716, + 0.013760631904006004, + 0.012252614833414555, + 0.007239635568112135, + 0.009169308468699455, + 0.012986350804567337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.020603211596608162 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.012547243386507034 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547243386507034, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547243386507034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24528855085372925, + 0.14857281744480133, + 0.07269047945737839, + 0.06565537303686142, + 0.04868321120738983, + 0.06336607784032822, + 0.0402534119784832, + 0.05095652863383293, + 0.11559996008872986, + 0.14893358945846558 + ], + "credit_map": [ + 0.0328608974814415, + 0.04421016573905945, + 0.02854539453983307, + 0.027728136628866196, + 0.04063979908823967, + 0.021198086440563202, + 0.025449801236391068, + 0.024656275287270546, + 0.013110918924212456, + 0.022292235866189003, + 0.021121492609381676, + 0.01687944307923317, + 0.02961697243154049, + 0.019363701343536377, + 0.012824389152228832, + 0.013363664038479328, + 0.006398387718945742, + 0.01081337034702301, + 0.013582233339548111, + 0.0066937548108398914, + 0.013764388859272003, + 0.011292477138340473, + 0.0053496151231229305, + 0.0047962539829313755, + 0.01279777567833662, + 0.006941895931959152, + 0.01358025148510933, + 0.006762310862541199, + 0.007617623545229435, + 0.0063553983345627785, + 0.006661030929535627, + 0.004939086269587278, + 0.0072202240116894245, + 0.007205324713140726, + 0.006040574517101049, + 0.005974723491817713, + 0.006739673670381308, + 0.005238369107246399, + 0.005904851946979761, + 0.004359474405646324, + 0.008541081100702286, + 0.009407497011125088, + 0.008733690716326237, + 0.007159379776567221, + 0.008626437745988369, + 0.007206459529697895, + 0.007317270617932081, + 0.006374260410666466, + 0.00407980615273118, + 0.004150080028921366, + 0.0051232329569756985, + 0.008828726597130299, + 0.0043199555948376656, + 0.0028743508737534285, + 0.006661463063210249, + 0.004215796012431383, + 0.004695350304245949, + 0.006960700731724501, + 0.007019653916358948, + 0.008902502246201038, + 0.007668442092835903, + 0.006307860836386681, + 0.00471067288890481, + 0.004691344685852528, + 0.01729460060596466, + 0.026992132887244225, + 0.014423578046262264, + 0.01440481748431921, + 0.013054503127932549, + 0.011979231610894203, + 0.008715624921023846, + 0.008735469542443752, + 0.025111908093094826, + 0.011798533610999584, + 0.016062647104263306, + 0.03943543881177902, + 0.012280087918043137, + 0.01372633595019579, + 0.01562507636845112, + 0.014893567189574242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012547243386507034 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.694981098175049 + ], + "opposite_logprob_tokens": [ + -0.06998113542795181 + ], + "logprob_reward": -2.694981098175049, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.694981098175049, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1715242862701416, + 0.16782636940479279, + 0.08496783673763275, + 0.18139411509037018, + 0.05780314654111862, + 0.10409881174564362, + 0.05424284562468529, + 0.07931989431381226, + 0.04453704133629799, + 0.05428563058376312 + ], + "credit_map": [ + 0.02262370102107525, + 0.022775333374738693, + 0.03332319110631943, + 0.01899687573313713, + 0.02410542592406273, + 0.01675708219408989, + 0.015538389794528484, + 0.017404276877641678, + 0.015916207805275917, + 0.022634347900748253, + 0.027168232947587967, + 0.015665998682379723, + 0.031484272330999374, + 0.025366270914673805, + 0.013666972517967224, + 0.01592406816780567, + 0.00774609437212348, + 0.015151129104197025, + 0.011676513589918613, + 0.007035703398287296, + 0.020612608641386032, + 0.011370728723704815, + 0.006318526342511177, + 0.0050565279088914394, + 0.017176061868667603, + 0.0323689766228199, + 0.030374275520443916, + 0.01765274442732334, + 0.034525319933891296, + 0.026755962520837784, + 0.013794184662401676, + 0.008746576495468616, + 0.011933621019124985, + 0.00577356293797493, + 0.007425116840749979, + 0.006223510485142469, + 0.005891910754144192, + 0.00890257116407156, + 0.006535766646265984, + 0.00511708902195096, + 0.017712878063321114, + 0.019290462136268616, + 0.014867483638226986, + 0.010500259697437286, + 0.014297808520495892, + 0.008487523533403873, + 0.010947232134640217, + 0.00799515563994646, + 0.005498930811882019, + 0.008014358580112457, + 0.00790991447865963, + 0.010440059937536716, + 0.00914097297936678, + 0.0036029841285198927, + 0.003982376307249069, + 0.005653244908899069, + 0.008639859966933727, + 0.014919665642082691, + 0.008844652213156223, + 0.011701353825628757, + 0.015134390443563461, + 0.008804368786513805, + 0.005817504599690437, + 0.005458098370581865, + 0.0063169850036501884, + 0.009036601521074772, + 0.005905729252845049, + 0.006391044706106186, + 0.004397371783852577, + 0.00477093318477273, + 0.00316299544647336, + 0.004555378574877977, + 0.00970390159636736, + 0.006301268003880978, + 0.007789308670908213, + 0.008354121819138527, + 0.006179802119731903, + 0.004749758169054985, + 0.0049340734258294106, + 0.006273400969803333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -2.694981098175049 + }, + { + "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 crane with perfect white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crane startled from a perch atop a gnarled, ancient pine tree with dark, cracked bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the crane's wings flap decisively downward, revealing black primary feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth 360-degree orbit around the crane in ultra-slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glistening water droplets visible erupting from the pine needles and refracting warm golden 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": [ + -0.038120824843645096 + ], + "opposite_logprob_tokens": [ + -3.288120746612549 + ], + "logprob_reward": -0.038120824843645096, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038120824843645096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20677168667316437, + 0.15379279851913452, + 0.07424913346767426, + 0.19280333817005157, + 0.09775624424219131, + 0.0832904800772667, + 0.03575417399406433, + 0.04074006900191307, + 0.04080403968691826, + 0.074038065969944 + ], + "credit_map": [ + 0.031042015179991722, + 0.022942857816815376, + 0.034377001225948334, + 0.02705005556344986, + 0.03121284581720829, + 0.020012151449918747, + 0.01984766498208046, + 0.02028709277510643, + 0.01472463645040989, + 0.02113417163491249, + 0.027104491367936134, + 0.013627268373966217, + 0.028357109054923058, + 0.022312838584184647, + 0.012553410604596138, + 0.013978877104818821, + 0.007900677621364594, + 0.006920308340340853, + 0.015551099553704262, + 0.007206143345683813, + 0.013127870857715607, + 0.010520297102630138, + 0.006661642342805862, + 0.0063610984943807125, + 0.019666913896799088, + 0.035406116396188736, + 0.033319778740406036, + 0.020083695650100708, + 0.03746461495757103, + 0.025149447843432426, + 0.012831869535148144, + 0.00888090766966343, + 0.014637505635619164, + 0.01009945198893547, + 0.011421287432312965, + 0.02043205313384533, + 0.00848304107785225, + 0.013180269859731197, + 0.013994446955621243, + 0.005508183967322111, + 0.012746994383633137, + 0.009400494396686554, + 0.009740508161485195, + 0.010230598971247673, + 0.011192417703568935, + 0.010794056579470634, + 0.008077130652964115, + 0.011108282022178173, + 0.004531937185674906, + 0.003748474642634392, + 0.004676265176385641, + 0.007346294820308685, + 0.005267581436783075, + 0.0032269961666315794, + 0.003535496536642313, + 0.0034211259335279465, + 0.004191156476736069, + 0.007007040549069643, + 0.005382074974477291, + 0.006088501773774624, + 0.0067330170422792435, + 0.005259337369352579, + 0.003133679972961545, + 0.0029452587477862835, + 0.00525329215452075, + 0.007695403881371021, + 0.005311061628162861, + 0.004508018493652344, + 0.005483541171997786, + 0.005113208666443825, + 0.0037720922846347094, + 0.003667418612167239, + 0.010524825192987919, + 0.006804163567721844, + 0.008308548480272293, + 0.015701772645115852, + 0.007805479224771261, + 0.009585145860910416, + 0.007510653231292963, + 0.007797473110258579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.038120824843645096 + } + ], + "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/0000060/group_2/index_tensor([2967, 2967])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_tensor([2967, 2967])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2c80a4b10552dff1a76bd8229a86ea962d6253fe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_tensor([2967, 2967])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "silent, cinematic, low-angle, wide-angle shot, held static, captures the colossal scale of a massive bronze statue of a forgotten king as it is slowly devoured by the earth. The scene is set in a desolate ancient plaza paved with vast, cracked stone slabs, under an overcast sky casting a flat, aimless light. The statue itself depicts a stern, bearded monarch clad in imperial armor, coated in a thick, weathered green patina, its surface richly textured and ancient. The catastrophe began when a deep earthquake shook the ground, and tiny cracks began to form a spiderweb across the plaza's stone floor at the statue's base, washing away dust and mortar. Then, with a muffled groan of tormenting metal and rock, one side of the massive sculpture's base gave way, and the entire statue began to sink inexorably into the cracked surface. Its once-proud, regal stance tilted at a sickeningly unnatural angle, the king's body sagging in defeat. Thick clouds of ancient dust erupt from the abyss, small pebbles and dark soil flow along the statue's weed-covered surface, creating a somber yet stunning atmosphere of final, irreversible decay.", + "seed": 24369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24369_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": -0.30628608018159864, + "vlm_logprob": -0.30628608018159864, + "vlm_yes_no_margin": 2.1000000119209288, + "vlm_token_credit": -0.30628608018159864, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.28116586804389954 + ], + "opposite_logprob_tokens": [ + -1.6561658382415771 + ], + "logprob_reward": -0.28116586804389954, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.28116586804389954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20814353227615356, + 0.09790760278701782, + 0.047864023596048355, + 0.03563845157623291, + 0.13277515769004822, + 0.035301435738801956, + 0.044348884373903275, + 0.16850358247756958, + 0.08911797404289246, + 0.14039933681488037 + ], + "credit_map": [ + 0.02937205880880356, + 0.025201715528964996, + 0.02542153373360634, + 0.02821621485054493, + 0.02167849987745285, + 0.018999073654413223, + 0.03169877454638481, + 0.02755565382540226, + 0.007736031897366047, + 0.012597908265888691, + 0.014011855237185955, + 0.012849063612520695, + 0.012399538420140743, + 0.008225694298744202, + 0.01792622171342373, + 0.012161292135715485, + 0.004089798778295517, + 0.005296044051647186, + 0.007591903675347567, + 0.00820181518793106, + 0.005353772081434727, + 0.004575508181005716, + 0.00597429508343339, + 0.00678089028224349, + 0.003664308227598667, + 0.004917938262224197, + 0.003813298186287284, + 0.005849702749401331, + 0.003659615758806467, + 0.0038725100457668304, + 0.0053428844548761845, + 0.004518195055425167, + 0.012781611643731594, + 0.023017222061753273, + 0.017646005377173424, + 0.021044665947556496, + 0.013515859842300415, + 0.011977226473391056, + 0.01718142069876194, + 0.01561114564538002, + 0.0027363367844372988, + 0.004978175275027752, + 0.004086978733539581, + 0.005810513626784086, + 0.003924839664250612, + 0.002712100278586149, + 0.00705543952062726, + 0.003997051157057285, + 0.004173487424850464, + 0.006637671031057835, + 0.007863233797252178, + 0.009211554192006588, + 0.003695087041705847, + 0.0033968740608543158, + 0.005695479456335306, + 0.003675497369840741, + 0.013166961260139942, + 0.03262513503432274, + 0.01564778760075569, + 0.02328876592218876, + 0.02397007867693901, + 0.016575131565332413, + 0.02506091073155403, + 0.01816880702972412, + 0.008314463309943676, + 0.012723461724817753, + 0.013271395117044449, + 0.0153580317273736, + 0.012629668228328228, + 0.008216727524995804, + 0.009500617161393166, + 0.009103605523705482, + 0.012905453331768513, + 0.014799614436924458, + 0.02266009710729122, + 0.03431713953614235, + 0.010267353616654873, + 0.015139631927013397, + 0.016633858904242516, + 0.013676177710294724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.28116586804389954 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371135711669922 + ], + "opposite_logprob_tokens": [ + -0.38711363077163696 + ], + "logprob_reward": -1.1371135711669922, + "yes_no_margin_reward": -0.7499999403953552, + "token_credit_reward": -1.1371135711669922, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08197589218616486, + 0.05009213835000992, + 0.10410458594560623, + 0.07085507363080978, + 0.08382263779640198, + 0.07847467064857483, + 0.17417548596858978, + 0.1252574771642685, + 0.07971769571304321, + 0.15152430534362793 + ], + "credit_map": [ + 0.007791280280798674, + 0.009986589662730694, + 0.01011622790247202, + 0.013202252797782421, + 0.01059224084019661, + 0.007374519947916269, + 0.01272528525441885, + 0.010187499225139618, + 0.005345137324184179, + 0.006592472083866596, + 0.00687018595635891, + 0.007343474309891462, + 0.005563964601606131, + 0.0047572217881679535, + 0.0074042342603206635, + 0.006215449422597885, + 0.00905743520706892, + 0.010175482369959354, + 0.01676180586218834, + 0.014802748337388039, + 0.01133956853300333, + 0.008926310576498508, + 0.01819528080523014, + 0.014845963567495346, + 0.0071063051000237465, + 0.007524226792156696, + 0.009593408554792404, + 0.010400406084954739, + 0.006841736845672131, + 0.00837273895740509, + 0.0122364591807127, + 0.008779797703027725, + 0.008747529238462448, + 0.01632685400545597, + 0.01025440264493227, + 0.014106690883636475, + 0.007080490235239267, + 0.009377834387123585, + 0.010878036729991436, + 0.007050808984786272, + 0.0073412456549704075, + 0.01629682444036007, + 0.0075657907873392105, + 0.012039728462696075, + 0.006680972408503294, + 0.008061828091740608, + 0.012388310395181179, + 0.008099975064396858, + 0.018335968255996704, + 0.026092009618878365, + 0.0247505996376276, + 0.03868722543120384, + 0.01198483444750309, + 0.015859832987189293, + 0.023542454466223717, + 0.014922584407031536, + 0.015525217168033123, + 0.011513198725879192, + 0.028917090967297554, + 0.01977866142988205, + 0.01056203804910183, + 0.011405195109546185, + 0.016210665926337242, + 0.011345424689352512, + 0.007766611408442259, + 0.012696895748376846, + 0.008723689243197441, + 0.012812867760658264, + 0.009937926195561886, + 0.008094031363725662, + 0.010693325661122799, + 0.008992355316877365, + 0.015148034319281578, + 0.01865185983479023, + 0.024783113971352577, + 0.03445900231599808, + 0.012099281884729862, + 0.015454337932169437, + 0.019112303853034973, + 0.011816389858722687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -1.1371135711669922 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.048683278262615204 + ], + "opposite_logprob_tokens": [ + -3.0486831665039062 + ], + "logprob_reward": -0.048683278262615204, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683278262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12685877084732056, + 0.12953326106071472, + 0.10875918716192245, + 0.035523321479558945, + 0.28778064250946045, + 0.036669474095106125, + 0.03375441953539848, + 0.046140026301145554, + 0.16466455161571503, + 0.030316291376948357 + ], + "credit_map": [ + 0.011532396078109741, + 0.016821594908833504, + 0.02040107734501362, + 0.01583302766084671, + 0.015401027165353298, + 0.010077225975692272, + 0.01962685026228428, + 0.017165588214993477, + 0.01295444741845131, + 0.015610718168318272, + 0.023007972165942192, + 0.018277762457728386, + 0.015883484855294228, + 0.011176478117704391, + 0.01805933751165867, + 0.014563078992068768, + 0.010961835272610188, + 0.009399848990142345, + 0.019594471901655197, + 0.01322171650826931, + 0.012789235450327396, + 0.009893692098557949, + 0.014492231421172619, + 0.018406163901090622, + 0.003726040478795767, + 0.005441523157060146, + 0.006044378504157066, + 0.005124152172356844, + 0.0033807270228862762, + 0.003153521101921797, + 0.0038107349537312984, + 0.004842246416956186, + 0.02524695359170437, + 0.04463649541139603, + 0.02926737070083618, + 0.04864760488271713, + 0.023187922313809395, + 0.03836771473288536, + 0.04860991612076759, + 0.02981671504676342, + 0.004111416172236204, + 0.005739386659115553, + 0.005201365798711777, + 0.005187020171433687, + 0.003319019917398691, + 0.003418569453060627, + 0.00564134307205677, + 0.004051359370350838, + 0.003441859735175967, + 0.0035291663371026516, + 0.004767337813973427, + 0.008317775093019009, + 0.0019068842520937324, + 0.0038931393064558506, + 0.005629950203001499, + 0.0022683122660964727, + 0.0037779679987579584, + 0.006058983970433474, + 0.007533783558756113, + 0.007418863475322723, + 0.004291478078812361, + 0.0038702699821442366, + 0.007931477390229702, + 0.0052572088316082954, + 0.015539459884166718, + 0.02005637250840664, + 0.02424115128815174, + 0.030746834352612495, + 0.01897389441728592, + 0.017320124432444572, + 0.022329488769173622, + 0.015457241795957088, + 0.003212141105905175, + 0.0034895609132945538, + 0.0060043358244001865, + 0.004732568748295307, + 0.002141662873327732, + 0.003276795381680131, + 0.0028669172897934914, + 0.004592310171574354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.048683278262615204 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.03812357783317566 + ], + "opposite_logprob_tokens": [ + -3.288123607635498 + ], + "logprob_reward": -0.03812357783317566, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812357783317566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08865976333618164, + 0.07887964695692062, + 0.05257066711783409, + 0.0537344254553318, + 0.045129358768463135, + 0.05510236322879791, + 0.1661202311515808, + 0.18157027661800385, + 0.20445658266544342, + 0.07377669215202332 + ], + "credit_map": [ + 0.009359301999211311, + 0.008974679745733738, + 0.009451241232454777, + 0.015115424990653992, + 0.0076189301908016205, + 0.007550559472292662, + 0.018006406724452972, + 0.012583218514919281, + 0.006416663061827421, + 0.01002806518226862, + 0.011575905606150627, + 0.00997061189264059, + 0.010028195567429066, + 0.006709964945912361, + 0.014858498238027096, + 0.009291739203035831, + 0.0042504603043198586, + 0.006983412895351648, + 0.0071969772689044476, + 0.007258390076458454, + 0.006432571914047003, + 0.004380408674478531, + 0.010114431381225586, + 0.005954016465693712, + 0.004348074551671743, + 0.006618462968617678, + 0.008081638254225254, + 0.009482652880251408, + 0.004319450352340937, + 0.004478433169424534, + 0.010096810758113861, + 0.006308903452008963, + 0.004811821039766073, + 0.00662198755890131, + 0.0054862420074641705, + 0.006913038901984692, + 0.0037727877497673035, + 0.005573434755206108, + 0.007016430143266916, + 0.004933618474751711, + 0.005029658321291208, + 0.00742578087374568, + 0.009449883364140987, + 0.007623023819178343, + 0.0054301000200212, + 0.005318209994584322, + 0.010268270038068295, + 0.004557439591735601, + 0.010237589478492737, + 0.025681806728243828, + 0.017028074711561203, + 0.043636973947286606, + 0.012626291252672672, + 0.014682537876069546, + 0.02997605688869953, + 0.012250890024006367, + 0.01352130901068449, + 0.02797473594546318, + 0.017197338864207268, + 0.04295308142900467, + 0.0179281122982502, + 0.019342180341482162, + 0.027905462309718132, + 0.014748047105967999, + 0.019010048359632492, + 0.027104243636131287, + 0.028098121285438538, + 0.033579371869564056, + 0.02133491262793541, + 0.01821073703467846, + 0.023780012503266335, + 0.03333912789821625, + 0.0066901808604598045, + 0.008004670962691307, + 0.012327252887189388, + 0.016349967569112778, + 0.005628939718008041, + 0.0074833533726632595, + 0.01045235712081194, + 0.006839970126748085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.03812357783317566 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.02634410560131073 + ], + "opposite_logprob_tokens": [ + -3.651344060897827 + ], + "logprob_reward": -0.02634410560131073, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634410560131073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17265650629997253, + 0.07093217968940735, + 0.11258374154567719, + 0.01947297528386116, + 0.0524737723171711, + 0.05583921819925308, + 0.033823806792497635, + 0.08106029778718948, + 0.12893538177013397, + 0.272222101688385 + ], + "credit_map": [ + 0.021278804168105125, + 0.01990525983273983, + 0.02257436141371727, + 0.022837847471237183, + 0.01395308505743742, + 0.014349617063999176, + 0.026977116242051125, + 0.030780410394072533, + 0.006926610600203276, + 0.007850748486816883, + 0.009774881415069103, + 0.008665544912219048, + 0.006562627386301756, + 0.005991330370306969, + 0.011167027056217194, + 0.013993403874337673, + 0.013579689897596836, + 0.008514780551195145, + 0.02102472260594368, + 0.014208560809493065, + 0.010143010877072811, + 0.01103077083826065, + 0.014557509683072567, + 0.019524697214365005, + 0.002246264601126313, + 0.002533148741349578, + 0.0025115006137639284, + 0.002419295022264123, + 0.002638729987666011, + 0.0018227278487756848, + 0.00248556025326252, + 0.0028157480992376804, + 0.004946739412844181, + 0.006268330849707127, + 0.007506371475756168, + 0.0075409309938549995, + 0.005226692650467157, + 0.006947918329387903, + 0.007224512752145529, + 0.006812276318669319, + 0.00552776176482439, + 0.006040965206921101, + 0.01052169967442751, + 0.007680993061512709, + 0.004161141812801361, + 0.00801820307970047, + 0.007281394209712744, + 0.0066070654429495335, + 0.003667886834591627, + 0.004890470765531063, + 0.004485097248107195, + 0.00503883883357048, + 0.002627826761454344, + 0.005553700961172581, + 0.004478026647120714, + 0.003081956645473838, + 0.006625073961913586, + 0.01383338961750269, + 0.01023428700864315, + 0.012877503409981728, + 0.011131146922707558, + 0.007228336296975613, + 0.011881699785590172, + 0.007248864509165287, + 0.010776543989777565, + 0.015980372205376625, + 0.02012578584253788, + 0.022182850167155266, + 0.014527042396366596, + 0.011755053885281086, + 0.014171259477734566, + 0.01941647380590439, + 0.0279228538274765, + 0.025389473885297775, + 0.046029239892959595, + 0.06794697046279907, + 0.01716899871826172, + 0.03244335949420929, + 0.02901655249297619, + 0.026304638013243675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02634410560131073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "silent, cinematic, low-angle, wide-angle shot, held static, captures the colossal scale of a massive bronze statue of a forgotten king as it is slowly devoured by the earth. The scene is set in a desolate ancient plaza paved with vast, cracked stone slabs, under an overcast sky casting a flat, aimless light. The statue itself depicts a stern, bearded monarch clad in imperial armor, coated in a thick, weathered green patina, its surface richly textured and ancient. The catastrophe began when a deep earthquake shook the ground, and tiny cracks began to form a spiderweb across the plaza's stone floor at the statue's base, washing away dust and mortar. Then, with a muffled groan of tormenting metal and rock, one side of the massive sculpture's base gave way, and the entire statue began to sink inexorably into the cracked surface. Its once-proud, regal stance tilted at a sickeningly unnatural angle, the king's body sagging in defeat. Thick clouds of ancient dust erupt from the abyss, small pebbles and dark soil flow along the statue's weed-covered surface, creating a somber yet stunning atmosphere of final, irreversible decay.", + "seed": 24370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24370_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": -0.25381571985781193, + "vlm_logprob": -0.25381571985781193, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.25381571985781193, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.18289831280708313 + ], + "opposite_logprob_tokens": [ + -2.0578982830047607 + ], + "logprob_reward": -0.18289831280708313, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.18289831280708313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18073713779449463, + 0.038719404488801956, + 0.08821038156747818, + 0.19040776789188385, + 0.09583582729101181, + 0.041820887476205826, + 0.06797685474157333, + 0.16096949577331543, + 0.08752378076314926, + 0.04779845476150513 + ], + "credit_map": [ + 0.027162225916981697, + 0.01534300297498703, + 0.013630262576043606, + 0.011501827277243137, + 0.04778824374079704, + 0.03169747069478035, + 0.01399171631783247, + 0.019622381776571274, + 0.006610990967601538, + 0.005593743175268173, + 0.0027042971923947334, + 0.004617541097104549, + 0.004694996867328882, + 0.005133223719894886, + 0.005533027928322554, + 0.0038315814454108477, + 0.021643461659550667, + 0.012126320973038673, + 0.005166170187294483, + 0.011523949913680553, + 0.009983603842556477, + 0.011672956869006157, + 0.01093145553022623, + 0.0051624649204313755, + 0.05202432721853256, + 0.019378725439310074, + 0.01191973127424717, + 0.03174318000674248, + 0.011577345430850983, + 0.027856675907969475, + 0.024659043177962303, + 0.011248735710978508, + 0.021818052977323532, + 0.010414235293865204, + 0.011094511486589909, + 0.009455514140427113, + 0.0051857116632163525, + 0.013298919424414635, + 0.008333575911819935, + 0.016235312446951866, + 0.008895445615053177, + 0.003947045188397169, + 0.0068405792117118835, + 0.005033927038311958, + 0.0020262657199054956, + 0.005415115039795637, + 0.0034053868148475885, + 0.006257123313844204, + 0.017584148794412613, + 0.007006363943219185, + 0.008004353381693363, + 0.004725022707134485, + 0.005769033450633287, + 0.010445204563438892, + 0.005297964904457331, + 0.00914476253092289, + 0.03467566892504692, + 0.0203000046312809, + 0.016843197867274284, + 0.019194507971405983, + 0.016473468393087387, + 0.021764082834124565, + 0.015923984348773956, + 0.01579459011554718, + 0.01640096865594387, + 0.010021945461630821, + 0.011721109040081501, + 0.011829892173409462, + 0.006196761969476938, + 0.011374161578714848, + 0.00920920167118311, + 0.010769736021757126, + 0.01060521136969328, + 0.005311927292495966, + 0.004670093301683664, + 0.004509272985160351, + 0.0040320465341210365, + 0.007029311265796423, + 0.004703553393483162, + 0.006937036756426096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.18289831280708313 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826224684715271 + ], + "opposite_logprob_tokens": [ + -0.576224684715271 + ], + "logprob_reward": -0.826224684715271, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826224684715271, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0790107324719429, + 0.1233159527182579, + 0.11246198415756226, + 0.09184291958808899, + 0.052948903292417526, + 0.11841686069965363, + 0.07410131394863129, + 0.08946352452039719, + 0.2350742220878601, + 0.023363688960671425 + ], + "credit_map": [ + 0.011572149582207203, + 0.008616720326244831, + 0.006594866048544645, + 0.00626039644703269, + 0.018438300117850304, + 0.012862227857112885, + 0.007848166860640049, + 0.00681788707152009, + 0.012806626036763191, + 0.019039276987314224, + 0.009173904545605183, + 0.01827292889356613, + 0.0187856275588274, + 0.015225231647491455, + 0.02079581283032894, + 0.009216522797942162, + 0.01429789699614048, + 0.0167557243257761, + 0.00667237676680088, + 0.02482210099697113, + 0.008436821401119232, + 0.011146556586027145, + 0.020024698227643967, + 0.010305789299309254, + 0.015693482011556625, + 0.008772465400397778, + 0.007811008021235466, + 0.02160595916211605, + 0.005698245018720627, + 0.00970793329179287, + 0.016380522400140762, + 0.006173279602080584, + 0.01323726773262024, + 0.004668268840759993, + 0.005344711244106293, + 0.008122474886476994, + 0.003302663331851363, + 0.007622047793120146, + 0.005021621007472277, + 0.0056298356503248215, + 0.02155032381415367, + 0.009573379531502724, + 0.013457806780934334, + 0.012504790909588337, + 0.006426478736102581, + 0.01577572524547577, + 0.01227987464517355, + 0.026848455891013145, + 0.019534792751073837, + 0.0072798593901097775, + 0.006623332854360342, + 0.006053098477423191, + 0.005341794341802597, + 0.011125362478196621, + 0.007186541799455881, + 0.01095651276409626, + 0.013770395889878273, + 0.0073958043940365314, + 0.010852623730897903, + 0.010490435175597668, + 0.006605976726859808, + 0.011253930628299713, + 0.014018096029758453, + 0.015076241455972195, + 0.06268170475959778, + 0.019215824082493782, + 0.014042956754565239, + 0.02092665061354637, + 0.012899205088615417, + 0.04091691970825195, + 0.024780860170722008, + 0.039610039442777634, + 0.005103149451315403, + 0.00304313818924129, + 0.0020122905261814594, + 0.0026380964554846287, + 0.0024366541765630245, + 0.0033963362220674753, + 0.001894742134027183, + 0.0028392765671014786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.826224684715271 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.07898899912834167 + ], + "opposite_logprob_tokens": [ + -2.578989028930664 + ], + "logprob_reward": -0.07898899912834167, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898899912834167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1372992992401123, + 0.054145827889442444, + 0.11214665323495865, + 0.150697261095047, + 0.068353570997715, + 0.04177079349756241, + 0.07772701233625412, + 0.10297442972660065, + 0.21648064255714417, + 0.03840452432632446 + ], + "credit_map": [ + 0.014556434005498886, + 0.016128169372677803, + 0.01055910810828209, + 0.014866401441395283, + 0.029889311641454697, + 0.021593604236841202, + 0.01815027929842472, + 0.011555994860827923, + 0.008032138459384441, + 0.008162274025380611, + 0.004457541741430759, + 0.005416890140622854, + 0.008539522998034954, + 0.0073222629725933075, + 0.007848716340959072, + 0.004366482608020306, + 0.0174872949719429, + 0.02530093304812908, + 0.006567381788045168, + 0.01537358295172453, + 0.010845622979104519, + 0.012820008210837841, + 0.013775764033198357, + 0.009976065717637539, + 0.020464574918150902, + 0.01604098081588745, + 0.007561018690466881, + 0.021808907389640808, + 0.03532489761710167, + 0.020402587950229645, + 0.02042047120630741, + 0.008673815988004208, + 0.009990527294576168, + 0.008202833123505116, + 0.00584378931671381, + 0.018535230308771133, + 0.0038390236441046, + 0.0070068975910544395, + 0.00903968047350645, + 0.005895589478313923, + 0.007751097436994314, + 0.005582519806921482, + 0.004282976500689983, + 0.005301516968756914, + 0.0034042182378470898, + 0.005848312750458717, + 0.0041448697447776794, + 0.005455280654132366, + 0.024843398481607437, + 0.008124688640236855, + 0.00632241927087307, + 0.004906166810542345, + 0.004859957844018936, + 0.014292898587882519, + 0.006952815223485231, + 0.007424674462527037, + 0.014447372406721115, + 0.015445450320839882, + 0.010465696454048157, + 0.012156669981777668, + 0.008625508286058903, + 0.01462799683213234, + 0.014659377746284008, + 0.012546357698738575, + 0.034002453088760376, + 0.02521240897476673, + 0.02160046435892582, + 0.030166225507855415, + 0.022468233481049538, + 0.02759513445198536, + 0.029130451381206512, + 0.026305286213755608, + 0.009335976094007492, + 0.00448577431961894, + 0.0034442758187651634, + 0.003866762388497591, + 0.0029098461382091045, + 0.006055902224034071, + 0.0030415363144129515, + 0.00526445172727108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.07898899912834167 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.1603870987892151 + ], + "opposite_logprob_tokens": [ + -1.9103870391845703 + ], + "logprob_reward": -0.1603870987892151, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603870987892151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11227534711360931, + 0.04085009545087814, + 0.04865853488445282, + 0.15877322852611542, + 0.10133777558803558, + 0.04322946071624756, + 0.09108921885490417, + 0.21192869544029236, + 0.1345100998878479, + 0.05734759569168091 + ], + "credit_map": [ + 0.018723225221037865, + 0.011992688290774822, + 0.009555735625326633, + 0.014182858169078827, + 0.01749010756611824, + 0.015437315218150616, + 0.012917673215270042, + 0.01197573821991682, + 0.007072879001498222, + 0.004976177588105202, + 0.003356482135131955, + 0.005426027346402407, + 0.004395775962620974, + 0.0054671321995556355, + 0.005106496624648571, + 0.005049123894423246, + 0.005588909611105919, + 0.007616978604346514, + 0.0029952949844300747, + 0.010873186402022839, + 0.0034969053231179714, + 0.005210115108639002, + 0.008899244479835033, + 0.003977903630584478, + 0.027449730783700943, + 0.018417377024888992, + 0.00865126308053732, + 0.02303396724164486, + 0.03084600530564785, + 0.021903973072767258, + 0.01843671128153801, + 0.010034197010099888, + 0.01714896969497204, + 0.012518318369984627, + 0.014927349984645844, + 0.012422789819538593, + 0.0050049470737576485, + 0.012270932085812092, + 0.008649157360196114, + 0.018395312130451202, + 0.006465590558946133, + 0.003842995036393404, + 0.005853617563843727, + 0.005616488866508007, + 0.002218252047896385, + 0.005142378620803356, + 0.005215787794440985, + 0.008874352090060711, + 0.024455146864056587, + 0.01103698555380106, + 0.007542234379798174, + 0.009471489116549492, + 0.007483829278498888, + 0.014387283474206924, + 0.007718001492321491, + 0.00899424497038126, + 0.037201207131147385, + 0.024416916072368622, + 0.032009515911340714, + 0.02663424238562584, + 0.019582873210310936, + 0.028805725276470184, + 0.019086478278040886, + 0.024191729724407196, + 0.0256669819355011, + 0.01618381403386593, + 0.013847429305315018, + 0.01732621155679226, + 0.013929872773587704, + 0.018444331362843513, + 0.016281388700008392, + 0.012830064631998539, + 0.015234689228236675, + 0.00648172153159976, + 0.0039054094813764095, + 0.006193592678755522, + 0.005412701517343521, + 0.009965475648641586, + 0.004137745592743158, + 0.006016259081661701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.1603870987892151 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.02057950384914875 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.02057950384914875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057950384914875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09276901930570602, + 0.04410988837480545, + 0.03950933739542961, + 0.23545169830322266, + 0.04838766157627106, + 0.030896080657839775, + 0.06438986957073212, + 0.20313718914985657, + 0.10446862876415253, + 0.13688062131404877 + ], + "credit_map": [ + 0.018887987360358238, + 0.01002554502338171, + 0.006518505048006773, + 0.010744296945631504, + 0.015525775030255318, + 0.013538968749344349, + 0.009936336427927017, + 0.007591605186462402, + 0.005338937509804964, + 0.0067156520672142506, + 0.0034124930389225483, + 0.006482833065092564, + 0.005515028722584248, + 0.005321001634001732, + 0.007132604252547026, + 0.004191339947283268, + 0.007994132116436958, + 0.007103074807673693, + 0.0026932116597890854, + 0.005815749987959862, + 0.003117222338914871, + 0.004802928306162357, + 0.0051614223048090935, + 0.0028215984348207712, + 0.051397230476140976, + 0.021503526717424393, + 0.012546098791062832, + 0.035727061331272125, + 0.03741985559463501, + 0.03369028866291046, + 0.02930109202861786, + 0.013866541907191277, + 0.010397829115390778, + 0.006193080917000771, + 0.004494062624871731, + 0.006403731182217598, + 0.003255966817960143, + 0.006583645008504391, + 0.004097404889762402, + 0.006961936596781015, + 0.007237674668431282, + 0.0022780271247029305, + 0.0034768031910061836, + 0.0037989190313965082, + 0.00202284031547606, + 0.0046316590160131454, + 0.0031078357715159655, + 0.004342321306467056, + 0.008672094903886318, + 0.0062015363946557045, + 0.012263663113117218, + 0.006984897889196873, + 0.005223965737968683, + 0.007642088457942009, + 0.005913535598665476, + 0.011488089337944984, + 0.04216375574469566, + 0.028485404327511787, + 0.018911117687821388, + 0.018089711666107178, + 0.0199343953281641, + 0.028246045112609863, + 0.020152568817138672, + 0.027154186740517616, + 0.02785859815776348, + 0.009803070686757565, + 0.0057432763278484344, + 0.010846887715160847, + 0.008016088046133518, + 0.018579062074422836, + 0.009109624661505222, + 0.014512019231915474, + 0.03199024870991707, + 0.014192142523825169, + 0.011590152978897095, + 0.015391797758638859, + 0.010165303945541382, + 0.020478419959545135, + 0.011391044594347477, + 0.021681511774659157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.02057950384914875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "silent, cinematic, low-angle, wide-angle shot, held static, captures the colossal scale of a massive bronze statue of a forgotten king as it is slowly devoured by the earth. The scene is set in a desolate ancient plaza paved with vast, cracked stone slabs, under an overcast sky casting a flat, aimless light. The statue itself depicts a stern, bearded monarch clad in imperial armor, coated in a thick, weathered green patina, its surface richly textured and ancient. The catastrophe began when a deep earthquake shook the ground, and tiny cracks began to form a spiderweb across the plaza's stone floor at the statue's base, washing away dust and mortar. Then, with a muffled groan of tormenting metal and rock, one side of the massive sculpture's base gave way, and the entire statue began to sink inexorably into the cracked surface. Its once-proud, regal stance tilted at a sickeningly unnatural angle, the king's body sagging in defeat. Thick clouds of ancient dust erupt from the abyss, small pebbles and dark soil flow along the statue's weed-covered surface, creating a somber yet stunning atmosphere of final, irreversible decay.", + "seed": 24371, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24371_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": -0.32117375656962394, + "vlm_logprob": -0.32117375656962394, + "vlm_yes_no_margin": 1.725, + "vlm_token_credit": -0.32117375656962394, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.2691197693347931 + ], + "opposite_logprob_tokens": [ + -1.6441197395324707 + ], + "logprob_reward": -0.2691197693347931, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2691197693347931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35291898250579834, + 0.10032172501087189, + 0.036614470183849335, + 0.07530787587165833, + 0.029212959110736847, + 0.09539153426885605, + 0.11949455738067627, + 0.08448661863803864, + 0.049960605800151825, + 0.05629069358110428 + ], + "credit_map": [ + 0.0735238790512085, + 0.05481833219528198, + 0.04122145473957062, + 0.02887529693543911, + 0.032845981419086456, + 0.04548707231879234, + 0.037788812071084976, + 0.03835812956094742, + 0.02538224682211876, + 0.008455459959805012, + 0.013707304373383522, + 0.01026937086135149, + 0.0076420907862484455, + 0.016321608796715736, + 0.008588572032749653, + 0.009955063462257385, + 0.005379563197493553, + 0.004027704242616892, + 0.005296051502227783, + 0.003918344154953957, + 0.004052127245813608, + 0.005057668313384056, + 0.004402264952659607, + 0.0044807433150708675, + 0.014966015703976154, + 0.008486791513860226, + 0.008032464422285557, + 0.008226888254284859, + 0.007900784723460674, + 0.010747403837740421, + 0.006695906165987253, + 0.010251617059111595, + 0.006633186712861061, + 0.002674926770851016, + 0.002371459500864148, + 0.0025902921333909035, + 0.0024979617446660995, + 0.004695771727710962, + 0.00555263739079237, + 0.0021967210341244936, + 0.016332698985934258, + 0.0067489598877727985, + 0.014882047660648823, + 0.00902440957725048, + 0.006330129690468311, + 0.012936483137309551, + 0.009908054023981094, + 0.019228747114539146, + 0.024482548236846924, + 0.012132056057453156, + 0.016102762892842293, + 0.01244762260466814, + 0.007872174493968487, + 0.015699293464422226, + 0.013100316748023033, + 0.017657773569226265, + 0.017238277941942215, + 0.00998335238546133, + 0.005378494970500469, + 0.008015044033527374, + 0.006577035877853632, + 0.01974155381321907, + 0.009984738193452358, + 0.0075681135058403015, + 0.005803614389151335, + 0.005030801985412836, + 0.006918365601450205, + 0.007637823931872845, + 0.0044218734838068485, + 0.005977165419608355, + 0.007445550989359617, + 0.006725405342876911, + 0.010808110237121582, + 0.004342739004641771, + 0.009495972655713558, + 0.004063628148287535, + 0.004972272086888552, + 0.009784284047782421, + 0.006184670142829418, + 0.00663901399821043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2691197693347931 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539737939834595 + ], + "opposite_logprob_tokens": [ + -0.4289737939834595 + ], + "logprob_reward": -1.0539737939834595, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539737939834595, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05307910963892937, + 0.055066924542188644, + 0.03181079775094986, + 0.04320213571190834, + 0.028516335412859917, + 0.03013826347887516, + 0.07028557360172272, + 0.42658379673957825, + 0.061535127460956573, + 0.1997818946838379 + ], + "credit_map": [ + 0.008002780377864838, + 0.00564945675432682, + 0.007601098623126745, + 0.004911235999315977, + 0.0044966996647417545, + 0.007571661379188299, + 0.005370072089135647, + 0.00947610568255186, + 0.012589874677360058, + 0.0037094000726938248, + 0.008099649101495743, + 0.005313531961292028, + 0.003960757981985807, + 0.008281396701931953, + 0.005878077819943428, + 0.007234236691147089, + 0.0058838422410190105, + 0.0022176404017955065, + 0.006278547458350658, + 0.0034615020267665386, + 0.002477142494171858, + 0.003959862049669027, + 0.003361140377819538, + 0.004171121399849653, + 0.006966781336814165, + 0.004006922710686922, + 0.007316206581890583, + 0.005164452362805605, + 0.004250703379511833, + 0.00559780839830637, + 0.004418220836669207, + 0.005481040105223656, + 0.004718974232673645, + 0.004203248769044876, + 0.002514596562832594, + 0.002824025461450219, + 0.004161459859460592, + 0.0038669572677463293, + 0.003991839941591024, + 0.0022352354135364294, + 0.004998712334781885, + 0.0030940123833715916, + 0.003182405373081565, + 0.003017901908606291, + 0.0032528492156416178, + 0.004152259789407253, + 0.004270829726010561, + 0.00416929367929697, + 0.014848465099930763, + 0.00484226131811738, + 0.006850926671177149, + 0.0071379211731255054, + 0.005050871521234512, + 0.011138523928821087, + 0.008943713270127773, + 0.011472892947494984, + 0.15244928002357483, + 0.012972167693078518, + 0.01487970445305109, + 0.018629489466547966, + 0.0119934668764472, + 0.18049854040145874, + 0.019517932087183, + 0.015643225982785225, + 0.00865981075912714, + 0.004563474096357822, + 0.011788798496127129, + 0.007295815274119377, + 0.005725192837417126, + 0.008381463587284088, + 0.007055200636386871, + 0.008065368980169296, + 0.04985640197992325, + 0.013151620514690876, + 0.01437925361096859, + 0.013603250496089458, + 0.009544839151203632, + 0.06404325366020203, + 0.020622357726097107, + 0.014580920338630676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -1.0539737939834595 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.2015915811061859 + ], + "opposite_logprob_tokens": [ + -1.7015916109085083 + ], + "logprob_reward": -0.2015915811061859, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015915811061859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2199355959892273, + 0.07456494867801666, + 0.06869697570800781, + 0.06230829656124115, + 0.026434777304530144, + 0.05061976984143257, + 0.08503507077693939, + 0.2717966139316559, + 0.02986828237771988, + 0.11073967069387436 + ], + "credit_map": [ + 0.05005718767642975, + 0.028928590938448906, + 0.030287062749266624, + 0.016126487404108047, + 0.018866267055273056, + 0.03075277991592884, + 0.01875600963830948, + 0.026161188259720802, + 0.018437465652823448, + 0.006706629879772663, + 0.00899092573672533, + 0.007135927677154541, + 0.007992377504706383, + 0.010631945915520191, + 0.008468862622976303, + 0.006200806703418493, + 0.010949512012302876, + 0.007104615215212107, + 0.007745834998786449, + 0.007924727164208889, + 0.010666023008525372, + 0.009126529097557068, + 0.00845463015139103, + 0.006725101266056299, + 0.01343568041920662, + 0.005164110101759434, + 0.00734093738719821, + 0.005844138562679291, + 0.008356669917702675, + 0.009113818407058716, + 0.0059487889520823956, + 0.007104148622602224, + 0.004914354532957077, + 0.0034622889943420887, + 0.002862178022041917, + 0.0027225411031395197, + 0.002871053060516715, + 0.00329368538223207, + 0.0038313132245093584, + 0.0024773599579930305, + 0.009064337238669395, + 0.004806929267942905, + 0.005494273267686367, + 0.006254471838474274, + 0.004397185519337654, + 0.007409563288092613, + 0.008318969048559666, + 0.0048740338534116745, + 0.014079811982810497, + 0.007781614549458027, + 0.012772832997143269, + 0.010943364351987839, + 0.004890172742307186, + 0.00942455604672432, + 0.011142483912408352, + 0.014000235125422478, + 0.09187427908182144, + 0.011127496138215065, + 0.012193522416055202, + 0.012783982791006565, + 0.010772721841931343, + 0.10563955456018448, + 0.017209231853485107, + 0.010195814073085785, + 0.0038660336285829544, + 0.0035027291160076857, + 0.003745844354853034, + 0.0034673279151320457, + 0.0023990017361938953, + 0.0057465024292469025, + 0.00392451137304306, + 0.003216329962015152, + 0.020923951640725136, + 0.008586248382925987, + 0.009914573282003403, + 0.007586618419736624, + 0.008972810581326485, + 0.025613820180296898, + 0.01889675483107567, + 0.010244886390864849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.2015915811061859 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.04308943450450897 + ], + "opposite_logprob_tokens": [ + -3.1680893898010254 + ], + "logprob_reward": -0.04308943450450897, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308943450450897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34458377957344055, + 0.06915701925754547, + 0.054485078901052475, + 0.048058319836854935, + 0.0754145160317421, + 0.08869113028049469, + 0.04403429850935936, + 0.1436927169561386, + 0.03819941356778145, + 0.09368380904197693 + ], + "credit_map": [ + 0.06654490530490875, + 0.05119636282324791, + 0.04136793687939644, + 0.029399840161204338, + 0.031767576932907104, + 0.04467625543475151, + 0.03930184990167618, + 0.04032905027270317, + 0.01801859773695469, + 0.004517091903835535, + 0.008614348247647285, + 0.006585586816072464, + 0.007548393681645393, + 0.010646646842360497, + 0.006127959582954645, + 0.007098397705703974, + 0.015327735804021358, + 0.0035024851094931364, + 0.008538908325135708, + 0.005445160903036594, + 0.0037262788973748684, + 0.0080960001796484, + 0.00429987907409668, + 0.0055486285127699375, + 0.009077067486941814, + 0.00361974211409688, + 0.007661655079573393, + 0.004521683789789677, + 0.0036908567417412996, + 0.007556254044175148, + 0.004187050275504589, + 0.00774400494992733, + 0.014424464665353298, + 0.0058446722105145454, + 0.01142682135105133, + 0.007589366286993027, + 0.005289354827255011, + 0.010621755383908749, + 0.00836216565221548, + 0.011855904944241047, + 0.012588650919497013, + 0.007867967709898949, + 0.012904982082545757, + 0.008920310065150261, + 0.0070194979198277, + 0.010541039519011974, + 0.010338114574551582, + 0.01851055771112442, + 0.006638855207711458, + 0.0046132695861160755, + 0.0058623082004487514, + 0.0057030050083994865, + 0.0038612631615251303, + 0.005892036482691765, + 0.005872178822755814, + 0.005591378081589937, + 0.03939170390367508, + 0.009188112802803516, + 0.008848140016198158, + 0.011107662692666054, + 0.007408862467855215, + 0.04394831508398056, + 0.012215639464557171, + 0.0115842679515481, + 0.004200160037726164, + 0.004284210968762636, + 0.004081096034497023, + 0.005077447276562452, + 0.005130581092089415, + 0.005504598841071129, + 0.006097009871155024, + 0.0038243047893047333, + 0.013368472456932068, + 0.009871063753962517, + 0.012479131110012531, + 0.011495903134346008, + 0.00967446994036436, + 0.01408624742180109, + 0.012027720920741558, + 0.010680796578526497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.04308943450450897 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.03809420391917229 + ], + "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": [ + 0.2020493745803833, + 0.18806684017181396, + 0.0729207769036293, + 0.14543208479881287, + 0.027273379266262054, + 0.033873703330755234, + 0.1053648367524147, + 0.05328124761581421, + 0.11324106901884079, + 0.058496639132499695 + ], + "credit_map": [ + 0.045074108988046646, + 0.025800833478569984, + 0.028540628030896187, + 0.016408903524279594, + 0.017822617664933205, + 0.02819540537893772, + 0.01753605529665947, + 0.022670846432447433, + 0.0513957142829895, + 0.012575365602970123, + 0.03099839761853218, + 0.014199969358742237, + 0.015404078178107738, + 0.03086218610405922, + 0.015539932064712048, + 0.01709119975566864, + 0.013190106488764286, + 0.006675082724541426, + 0.013914930634200573, + 0.007493022363632917, + 0.005709742661565542, + 0.010877814143896103, + 0.0060899946838617325, + 0.008970090188086033, + 0.03360694274306297, + 0.010781053453683853, + 0.020512772724032402, + 0.01522815227508545, + 0.011626536026597023, + 0.021317223086953163, + 0.014290009625256062, + 0.01806940883398056, + 0.004554539453238249, + 0.0032555742654949427, + 0.003310831030830741, + 0.0027700045611709356, + 0.0029441670048981905, + 0.0036749006249010563, + 0.0038431754801422358, + 0.0029201884754002094, + 0.006964940577745438, + 0.0032669994980096817, + 0.0047316099517047405, + 0.0029417616315186024, + 0.0024214235600084066, + 0.0045686690136790276, + 0.003662002505734563, + 0.005316297058016062, + 0.02508624829351902, + 0.008274782449007034, + 0.012167362496256828, + 0.013080778531730175, + 0.007491962984204292, + 0.015177922323346138, + 0.011619162745773792, + 0.012466629035770893, + 0.01503552496433258, + 0.002558954758569598, + 0.0030521333683282137, + 0.004612816032022238, + 0.002377598313614726, + 0.017620857805013657, + 0.0041582281701266766, + 0.0038651402574032545, + 0.01442340761423111, + 0.00938415341079235, + 0.014194340445101261, + 0.014605413191020489, + 0.01167271938174963, + 0.01825234293937683, + 0.016264740377664566, + 0.01444395910948515, + 0.01495145633816719, + 0.003653045743703842, + 0.004125501494854689, + 0.00439264252781868, + 0.003193796146661043, + 0.019085735082626343, + 0.005278468132019043, + 0.003815997391939163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03809420391917229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "silent, cinematic, low-angle, wide-angle shot, held static, captures the colossal scale of a massive bronze statue of a forgotten king as it is slowly devoured by the earth. The scene is set in a desolate ancient plaza paved with vast, cracked stone slabs, under an overcast sky casting a flat, aimless light. The statue itself depicts a stern, bearded monarch clad in imperial armor, coated in a thick, weathered green patina, its surface richly textured and ancient. The catastrophe began when a deep earthquake shook the ground, and tiny cracks began to form a spiderweb across the plaza's stone floor at the statue's base, washing away dust and mortar. Then, with a muffled groan of tormenting metal and rock, one side of the massive sculpture's base gave way, and the entire statue began to sink inexorably into the cracked surface. Its once-proud, regal stance tilted at a sickeningly unnatural angle, the king's body sagging in defeat. Thick clouds of ancient dust erupt from the abyss, small pebbles and dark soil flow along the statue's weed-covered surface, creating a somber yet stunning atmosphere of final, irreversible decay.", + "seed": 24372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_2/index_2967_seed_24372_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": -0.30004370473325254, + "vlm_logprob": -0.30004370473325254, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.30004370473325254, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.17443779110908508 + ], + "opposite_logprob_tokens": [ + -2.0494377613067627 + ], + "logprob_reward": -0.17443779110908508, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17443779110908508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14929336309432983, + 0.04698329046368599, + 0.038951680064201355, + 0.10332204401493073, + 0.03286068141460419, + 0.02369273453950882, + 0.0763658955693245, + 0.3106876015663147, + 0.10567236691713333, + 0.11217031627893448 + ], + "credit_map": [ + 0.014525139704346657, + 0.012023967690765858, + 0.01896018534898758, + 0.022280069068074226, + 0.022465428337454796, + 0.016423815861344337, + 0.021327819675207138, + 0.02128693088889122, + 0.005199003033339977, + 0.006317756604403257, + 0.005240365397185087, + 0.007529157213866711, + 0.006558659486472607, + 0.004783558193594217, + 0.007196967490017414, + 0.004157823510468006, + 0.005105610936880112, + 0.004400892183184624, + 0.00513515155762434, + 0.0041742646135389805, + 0.0041502052918076515, + 0.004766370169818401, + 0.005745230242609978, + 0.005473952740430832, + 0.018727228045463562, + 0.012694942764937878, + 0.015935860574245453, + 0.010871958918869495, + 0.005923692602664232, + 0.014147807843983173, + 0.011913410387933254, + 0.013107141479849815, + 0.003961543086916208, + 0.0027528919745236635, + 0.005864781327545643, + 0.0035693147219717503, + 0.003306074533611536, + 0.0027900789864361286, + 0.004100480582565069, + 0.006515517830848694, + 0.002760767238214612, + 0.004180669318884611, + 0.002744527766481042, + 0.0025460110045969486, + 0.003636883106082678, + 0.0014464721316471696, + 0.004174154717475176, + 0.002203248906880617, + 0.0061278678476810455, + 0.006748814135789871, + 0.007979062385857105, + 0.018710942938923836, + 0.00809495523571968, + 0.006009806413203478, + 0.011685701087117195, + 0.011008745059370995, + 0.013544300571084023, + 0.015264498069882393, + 0.02947421185672283, + 0.13118106126785278, + 0.012786710634827614, + 0.008311049081385136, + 0.08886769413948059, + 0.011258081533014774, + 0.005033675115555525, + 0.005988454446196556, + 0.012434613890945911, + 0.03973161056637764, + 0.00548430485650897, + 0.004448292311280966, + 0.025757968425750732, + 0.0067934514954686165, + 0.007303163409233093, + 0.012424921616911888, + 0.021244989708065987, + 0.018368497490882874, + 0.01242363266646862, + 0.007243522442877293, + 0.022101037204265594, + 0.011060556396842003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.17443779110908508 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371376514434814 + ], + "opposite_logprob_tokens": [ + -0.38713768124580383 + ], + "logprob_reward": -1.1371376514434814, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371376514434814, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.016882073134183884, + 0.022747136652469635, + 0.05425325781106949, + 0.04973346367478371, + 0.036360204219818115, + 0.03528176620602608, + 0.030465055257081985, + 0.6150111556053162, + 0.09455610811710358, + 0.04470977559685707 + ], + "credit_map": [ + 0.0016564085381105542, + 0.001466840854845941, + 0.0015057515120133758, + 0.0031988052651286125, + 0.0024974383413791656, + 0.0017611568327993155, + 0.0033547901548445225, + 0.0014408818678930402, + 0.002170952968299389, + 0.0033408361487090588, + 0.0019532099831849337, + 0.003249553730711341, + 0.003351819235831499, + 0.0024884268641471863, + 0.004399523138999939, + 0.0017928141169250011, + 0.006830307189375162, + 0.0077513838186860085, + 0.008180679753422737, + 0.005197945050895214, + 0.004601022228598595, + 0.005722563248127699, + 0.007743456866592169, + 0.008225901052355766, + 0.00827021524310112, + 0.004327966831624508, + 0.008303770795464516, + 0.0062194461934268475, + 0.003995315637439489, + 0.006313977297395468, + 0.005846330430358648, + 0.006456437986344099, + 0.0031947968527674675, + 0.0033229244872927666, + 0.0058824652805924416, + 0.006853744853287935, + 0.003944781143218279, + 0.0023388657718896866, + 0.006727868225425482, + 0.004094753414392471, + 0.004649932496249676, + 0.0031586582772433758, + 0.00381700717844069, + 0.0044699194841086864, + 0.00480642169713974, + 0.0038917374331504107, + 0.005923785734921694, + 0.004564302507787943, + 0.0025766948238015175, + 0.0038296133279800415, + 0.0032484408002346754, + 0.0034530996344983578, + 0.00452663516625762, + 0.0029048011638224125, + 0.004892909899353981, + 0.005032859742641449, + 0.025428729131817818, + 0.023620102554559708, + 0.05783191695809364, + 0.27226755023002625, + 0.015172199346125126, + 0.015433147549629211, + 0.18363766372203827, + 0.021619856357574463, + 0.0049788993783295155, + 0.006348957773298025, + 0.013936578296124935, + 0.028264155611395836, + 0.009431452490389347, + 0.005129094235599041, + 0.01933390274643898, + 0.007133073639124632, + 0.002490825019776821, + 0.003508243476971984, + 0.008123083971440792, + 0.00995250977575779, + 0.006080588325858116, + 0.0021415380761027336, + 0.008571980521082878, + 0.003841004567220807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -1.1371376514434814 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.08906017988920212 + ], + "opposite_logprob_tokens": [ + -2.464060068130493 + ], + "logprob_reward": -0.08906017988920212, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906017988920212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09282300621271133, + 0.07959578931331635, + 0.11785247921943665, + 0.11550473421812057, + 0.04787382856011391, + 0.06729717552661896, + 0.02969861775636673, + 0.30626216530799866, + 0.11499562859535217, + 0.02809656225144863 + ], + "credit_map": [ + 0.012025442905724049, + 0.007810191251337528, + 0.011361447162926197, + 0.015733055770397186, + 0.011458945460617542, + 0.00898499321192503, + 0.015367604792118073, + 0.010081324726343155, + 0.0077849519439041615, + 0.014132310636341572, + 0.006073311436921358, + 0.010990971699357033, + 0.012671386823058128, + 0.008920653723180294, + 0.012806770391762257, + 0.006215436849743128, + 0.01300309132784605, + 0.020088618621230125, + 0.013315864838659763, + 0.01601950079202652, + 0.01214295532554388, + 0.012364986352622509, + 0.020654946565628052, + 0.01026251632720232, + 0.017643427476286888, + 0.012009608559310436, + 0.0154399573802948, + 0.015066597610712051, + 0.012246925383806229, + 0.013125834055244923, + 0.015169072896242142, + 0.014803308993577957, + 0.004638324026018381, + 0.0036228441167622805, + 0.006980400066822767, + 0.007634163834154606, + 0.006013122852891684, + 0.0033154187258332968, + 0.008366281166672707, + 0.007303273770958185, + 0.007711108773946762, + 0.007589814253151417, + 0.008342036977410316, + 0.00985241960734129, + 0.009676726534962654, + 0.00490267388522625, + 0.011030271649360657, + 0.008192126639187336, + 0.0019005857175216079, + 0.003145369002595544, + 0.003741262946277857, + 0.006949721835553646, + 0.0029671567026525736, + 0.0018627673853188753, + 0.005682040937244892, + 0.003449713811278343, + 0.012771680019795895, + 0.013074892573058605, + 0.02871800772845745, + 0.13227961957454681, + 0.011354987509548664, + 0.007972309365868568, + 0.08889508247375488, + 0.011195586062967777, + 0.006642437539994717, + 0.008549845777451992, + 0.014331339858472347, + 0.0364568829536438, + 0.01209061499685049, + 0.0056001064367592335, + 0.024910591542720795, + 0.006413801573216915, + 0.0017248740186914802, + 0.003280719742178917, + 0.0037926791701465845, + 0.004042979795485735, + 0.004641852807253599, + 0.0020853509195148945, + 0.006312558893114328, + 0.002215546788647771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.08906017988920212 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.07900211215019226 + ], + "opposite_logprob_tokens": [ + -2.5790021419525146 + ], + "logprob_reward": -0.07900211215019226, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900211215019226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13574445247650146, + 0.09792813658714294, + 0.10091462731361389, + 0.09739616513252258, + 0.021410992369055748, + 0.03962417319417, + 0.05931625887751579, + 0.1994488537311554, + 0.19961042702198029, + 0.04860588163137436 + ], + "credit_map": [ + 0.015305709093809128, + 0.012031802907586098, + 0.016224442049860954, + 0.021822405979037285, + 0.01765335351228714, + 0.01629689894616604, + 0.020430119708180428, + 0.015979737043380737, + 0.008532779291272163, + 0.01283600740134716, + 0.00978965125977993, + 0.016878413036465645, + 0.013345952145755291, + 0.009385027922689915, + 0.018539031967520714, + 0.008621285669505596, + 0.011725706048309803, + 0.015989530831575394, + 0.01477771531790495, + 0.010863540694117546, + 0.009843800216913223, + 0.010538032278418541, + 0.013848748989403248, + 0.013327562250196934, + 0.016285836696624756, + 0.010828052647411823, + 0.013182044960558414, + 0.01249772123992443, + 0.009593099355697632, + 0.009789373725652695, + 0.012239384464919567, + 0.012980658560991287, + 0.001891432679258287, + 0.0016283869044855237, + 0.004095255397260189, + 0.0034339004196226597, + 0.0025705730076879263, + 0.0014339196495711803, + 0.003913156222552061, + 0.0024443708825856447, + 0.0043257554061710835, + 0.003409945871680975, + 0.00824450422078371, + 0.004055693279951811, + 0.004274051636457443, + 0.004709337372332811, + 0.005859819706529379, + 0.004745070822536945, + 0.0038888377603143454, + 0.006501704454421997, + 0.006223087664693594, + 0.013677561655640602, + 0.007674628868699074, + 0.0037862127646803856, + 0.010049662552773952, + 0.007514567580074072, + 0.008550490252673626, + 0.010455721989274025, + 0.01858387142419815, + 0.08269485831260681, + 0.009042897261679173, + 0.005290251690894365, + 0.05668835714459419, + 0.00814241822808981, + 0.008102383464574814, + 0.013298959471285343, + 0.02698047086596489, + 0.06600489467382431, + 0.01652122288942337, + 0.008051280863583088, + 0.0440894290804863, + 0.01656179502606392, + 0.002969487337395549, + 0.0035555399954319, + 0.008310381323099136, + 0.013652636669576168, + 0.004846779629588127, + 0.002676646690815687, + 0.00804247334599495, + 0.004551939200609922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.07900211215019226 + }, + { + "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\": \"Is the camera held static throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main subject a massive bronze statue of a bearded monarch wearing imperial armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the statue visibly sinking into the ground with a tilted, unnatural angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a desolate ancient plaza with vast, cracked stone slabs under an overcast 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": [ + -0.02058078907430172 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.02058078907430172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058078907430172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11795193701982498, + 0.15454131364822388, + 0.18663473427295685, + 0.028775570914149284, + 0.02128901705145836, + 0.018138226121664047, + 0.05239871144294739, + 0.3275866210460663, + 0.06166015937924385, + 0.031023692339658737 + ], + "credit_map": [ + 0.015627063810825348, + 0.009255606681108475, + 0.014890092425048351, + 0.016703950241208076, + 0.013780741952359676, + 0.013400656171143055, + 0.015629470348358154, + 0.018664352595806122, + 0.013478938490152359, + 0.02008867636322975, + 0.019379163160920143, + 0.027054401114583015, + 0.019885938614606857, + 0.014121418818831444, + 0.02326812408864498, + 0.017264647409319878, + 0.01707623526453972, + 0.031819071620702744, + 0.029033362865447998, + 0.02198304608464241, + 0.016607243567705154, + 0.018555982038378716, + 0.02778353914618492, + 0.02377624623477459, + 0.005274035502225161, + 0.002409188076853752, + 0.004133356735110283, + 0.0031698821112513542, + 0.0028817711863666773, + 0.0043347799219191074, + 0.003247222863137722, + 0.0033253333531320095, + 0.0016588526777923107, + 0.0017645582556724548, + 0.0032427830155938864, + 0.002937408396974206, + 0.003224091837182641, + 0.0014134328812360764, + 0.0029659688007086515, + 0.004081920720636845, + 0.0028210580348968506, + 0.0015496899140998721, + 0.002497292123734951, + 0.002072493080049753, + 0.002500998554751277, + 0.0019283141009509563, + 0.003136624349281192, + 0.0016317558474838734, + 0.003713191021233797, + 0.0032585151493549347, + 0.006130464840680361, + 0.016087790951132774, + 0.004302034620195627, + 0.0031440139282494783, + 0.010669931769371033, + 0.0050927684642374516, + 0.013964754529297352, + 0.016833936795592308, + 0.03270421922206879, + 0.12742671370506287, + 0.02265343628823757, + 0.012839024886488914, + 0.08608310669660568, + 0.015081401914358139, + 0.003480621613562107, + 0.004215396475046873, + 0.008224227465689182, + 0.017956390976905823, + 0.006552991922944784, + 0.0032728209625929594, + 0.012340214103460312, + 0.005617493763566017, + 0.0036323866806924343, + 0.00275125028565526, + 0.004021689761430025, + 0.004110547248274088, + 0.003982244990766048, + 0.004023184068500996, + 0.004724449478089809, + 0.0037779409904032946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.02058078907430172 + } + ], + "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/0000060/group_3/index_tensor([6239, 6239])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_tensor([6239, 6239])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..61f9c1da48cc7caefdcb59d1841dcaed3897f8be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_tensor([6239, 6239])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static close-up captures a dense field of slender, elongated fountain bamboo canes bracing themselves against a powerful gale. The entire scene is rendered in stunning ultra-high-definition slow motion, imbuing the natural event with the dramatic intensity of a high-speed sports documentary. Beneath a turbulent, dark-gray sky, each vibrantly green, segmented cane, its waxy surface gleaming, bends in a deep, tense arc, nearly driven to the ground by an unseen, overwhelming force. The camera maintains focus as the canes reach their elastic limit, drawing back with a violent, elastic recoil, their narrow leaves fluttering chaotically in a deep green blur. Dramatic, high-contrast light occasionally penetrates the fast-moving clouds, casting fleeting, sharp highlights on the tense stalks and individual water droplets flung from the leaves, freezing the moment of impact and release. The atmosphere is one of a fierce, primal struggle, showcasing the potent contest between the wind's brute force and the bamboo's profound flexibility.", + "seed": 28838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28838_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": -0.07814501971006393, + "vlm_logprob": -0.07814501971006393, + "vlm_yes_no_margin": 2.749999976158142, + "vlm_token_credit": -0.07814501971006393, + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06105884164571762 + ], + "opposite_logprob_tokens": [ + -3.311058759689331 + ], + "logprob_reward": -0.06105884164571762, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06105884164571762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023714950308203697, + 0.20202289521694183, + 0.0210706889629364, + 0.025148050859570503, + 0.027545079588890076, + 0.02265813574194908, + 0.03051936626434326, + 0.11772860586643219, + 0.463946133852005, + 0.06564611196517944 + ], + "credit_map": [ + 0.0033831752371042967, + 0.0027533848769962788, + 0.002621300285682082, + 0.0022919990587979555, + 0.002753244712948799, + 0.0037177633494138718, + 0.00262616784311831, + 0.003567915176972747, + 0.016482848674058914, + 0.009266704320907593, + 0.00957365334033966, + 0.005702636670321226, + 0.08136570453643799, + 0.05579699948430061, + 0.013054393231868744, + 0.01077994890511036, + 0.0026221938896924257, + 0.0015948619693517685, + 0.003592628752812743, + 0.0020215401891618967, + 0.002294570440426469, + 0.003694410203024745, + 0.0015883233863860369, + 0.003662160597741604, + 0.0022266460582613945, + 0.002561226487159729, + 0.00264430558308959, + 0.00185517652425915, + 0.0029467586427927017, + 0.004360490012913942, + 0.0032756600994616747, + 0.005277789197862148, + 0.0024466465692967176, + 0.002055578865110874, + 0.0016546306433156133, + 0.0016287406906485558, + 0.0019338532583788037, + 0.006146535277366638, + 0.0026770473923534155, + 0.00900204572826624, + 0.0027172730769962072, + 0.0023667383939027786, + 0.0023427747655659914, + 0.001871874206699431, + 0.001975481165573001, + 0.0036426724400371313, + 0.0025581575464457273, + 0.005183164495974779, + 0.004332211799919605, + 0.005008963868021965, + 0.003107577795162797, + 0.002183254575356841, + 0.0037449137307703495, + 0.00402969541028142, + 0.004049072507768869, + 0.00406367564573884, + 0.013529750518500805, + 0.01716453768312931, + 0.011189216747879982, + 0.010865340940654278, + 0.017005680128932, + 0.018162179738283157, + 0.014614107087254524, + 0.015197785571217537, + 0.07137937843799591, + 0.017307601869106293, + 0.020559754222631454, + 0.018870508298277855, + 0.019700655713677406, + 0.1658882349729538, + 0.016743242740631104, + 0.1334967464208603, + 0.006685656029731035, + 0.006225922144949436, + 0.00635678181424737, + 0.004087924025952816, + 0.005637678317725658, + 0.013395142741501331, + 0.003954631742089987, + 0.019302377477288246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.06105884164571762 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270943582057953 + ], + "opposite_logprob_tokens": [ + -2.127094268798828 + ], + "logprob_reward": -0.1270943582057953, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270943582057953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044928014278411865, + 0.1825796514749527, + 0.08578165620565414, + 0.06823916733264923, + 0.1430858075618744, + 0.028661945834755898, + 0.03265468776226044, + 0.06495802104473114, + 0.2827390134334564, + 0.06637202203273773 + ], + "credit_map": [ + 0.005390977021306753, + 0.0071770003996789455, + 0.0034394441172480583, + 0.004897608421742916, + 0.005725035443902016, + 0.004989481531083584, + 0.007560099940747023, + 0.005748366937041283, + 0.015168393962085247, + 0.011083745397627354, + 0.01246880367398262, + 0.007915004156529903, + 0.0648992732167244, + 0.04313267022371292, + 0.012741304002702236, + 0.015170465223491192, + 0.008586114272475243, + 0.00924389436841011, + 0.00874266866594553, + 0.007029587868601084, + 0.012327702715992928, + 0.01837628148496151, + 0.010991934686899185, + 0.010483467020094395, + 0.006391249597072601, + 0.007568288128823042, + 0.008741161786019802, + 0.005422606132924557, + 0.011046340689063072, + 0.009792777709662914, + 0.009827643632888794, + 0.009449095465242863, + 0.008308535441756248, + 0.008413298055529594, + 0.003901364980265498, + 0.00465491833165288, + 0.007597204763442278, + 0.050330325961112976, + 0.005975491367280483, + 0.05390466749668121, + 0.0031361363362520933, + 0.0038429356645792723, + 0.0028681145049631596, + 0.00270998477935791, + 0.003886164166033268, + 0.004309645388275385, + 0.004476075526326895, + 0.0034328897017985582, + 0.00461130915209651, + 0.00717100128531456, + 0.005307098384946585, + 0.0022095772437751293, + 0.003337044967338443, + 0.0030812020413577557, + 0.003333675442263484, + 0.0036037794779986143, + 0.007365940138697624, + 0.009382271207869053, + 0.006057127378880978, + 0.005915259476751089, + 0.009136960841715336, + 0.010484475642442703, + 0.008166277781128883, + 0.008449709974229336, + 0.0431375615298748, + 0.010578147135674953, + 0.012119989842176437, + 0.011973515152931213, + 0.012202057987451553, + 0.10092709958553314, + 0.010539750568568707, + 0.08126088976860046, + 0.01247368659824133, + 0.007618850562721491, + 0.00604461831972003, + 0.004350502043962479, + 0.0053795683197677135, + 0.015803806483745575, + 0.0034671658650040627, + 0.01123382244259119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.1270943582057953 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898811995983124 + ], + "opposite_logprob_tokens": [ + -2.5789880752563477 + ], + "logprob_reward": -0.07898811995983124, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898811995983124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03423961624503136, + 0.2641806900501251, + 0.05977236479520798, + 0.06165970489382744, + 0.30182531476020813, + 0.030808115378022194, + 0.045066140592098236, + 0.044581372290849686, + 0.10173564404249191, + 0.0561310350894928 + ], + "credit_map": [ + 0.004935227334499359, + 0.005255864467471838, + 0.0038186677265912294, + 0.003695073537528515, + 0.004452710039913654, + 0.003354918211698532, + 0.004563487600535154, + 0.004163671750575304, + 0.018112866207957268, + 0.01152826938778162, + 0.019221536815166473, + 0.008876417763531208, + 0.09948325157165527, + 0.0756978765130043, + 0.01578914187848568, + 0.01547135692089796, + 0.006218449678272009, + 0.005060494877398014, + 0.008607625961303711, + 0.005549045745283365, + 0.008232682943344116, + 0.012899335473775864, + 0.005495249293744564, + 0.0077094887383282185, + 0.006586839444935322, + 0.007922173477709293, + 0.008367432281374931, + 0.005394445266574621, + 0.009090024046599865, + 0.00818923581391573, + 0.008209598250687122, + 0.007899961434304714, + 0.013715877197682858, + 0.012509635649621487, + 0.007823379710316658, + 0.009944286197423935, + 0.01608067937195301, + 0.11067298799753189, + 0.01262820977717638, + 0.11845027655363083, + 0.002814103150740266, + 0.0030284805689007044, + 0.003530098358169198, + 0.0028677391819655895, + 0.004064342472702265, + 0.004365756642073393, + 0.005627480335533619, + 0.004510116297751665, + 0.00560715002939105, + 0.00521167041733861, + 0.0035364904906600714, + 0.0034743829164654016, + 0.007160916458815336, + 0.00429800059646368, + 0.009323079138994217, + 0.006454453803598881, + 0.005159763619303703, + 0.006624949164688587, + 0.004745034035295248, + 0.004075770732015371, + 0.0065507129766047, + 0.006372032221406698, + 0.005568747874349356, + 0.00548436539247632, + 0.012741961516439915, + 0.0055367411114275455, + 0.006997847929596901, + 0.007603785023093224, + 0.005809869151562452, + 0.03131483495235443, + 0.005601950921118259, + 0.02612866461277008, + 0.008938644081354141, + 0.007973622530698776, + 0.007506003137677908, + 0.004048195201903582, + 0.005106027238070965, + 0.01006005983799696, + 0.002932868432253599, + 0.009565617889165878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.07898811995983124 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328725904226303 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.02328725904226303, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328725904226303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026733212172985077, + 0.07488825172185898, + 0.040696777403354645, + 0.11419747024774551, + 0.08100996166467667, + 0.04164183884859085, + 0.02533768117427826, + 0.10094498842954636, + 0.4691339433193207, + 0.02541583962738514 + ], + "credit_map": [ + 0.004469801671802998, + 0.0032869891729205847, + 0.004823160357773304, + 0.002605827758088708, + 0.0030645164661109447, + 0.0028880215249955654, + 0.0028139755595475435, + 0.0027809233870357275, + 0.0077900756150484085, + 0.004181923344731331, + 0.004927601665258408, + 0.0035892585292458534, + 0.017805129289627075, + 0.020086148753762245, + 0.00619916245341301, + 0.010308966040611267, + 0.004488268867135048, + 0.003457113401964307, + 0.007526230067014694, + 0.0047222343273460865, + 0.005113687831908464, + 0.006605775095522404, + 0.0035041661467403173, + 0.005279308184981346, + 0.014935033395886421, + 0.018752802163362503, + 0.015605905093252659, + 0.009413192048668861, + 0.011837046593427658, + 0.01591425947844982, + 0.009501409716904163, + 0.018237827345728874, + 0.01270835381001234, + 0.008626981638371944, + 0.002836898434907198, + 0.0025794203393161297, + 0.005339676048606634, + 0.019185885787010193, + 0.003587606130167842, + 0.026145154610276222, + 0.004143466241657734, + 0.004097295459359884, + 0.0041811647824943066, + 0.0040130638517439365, + 0.003628627397119999, + 0.00634734146296978, + 0.0045034438371658325, + 0.010727438144385815, + 0.0035624741576611996, + 0.005007397383451462, + 0.00268742348998785, + 0.0018100108718499541, + 0.003136079292744398, + 0.0031767881009727716, + 0.003386989701539278, + 0.002570522716268897, + 0.01228110957890749, + 0.015168563462793827, + 0.00893345009535551, + 0.008731625974178314, + 0.014044550247490406, + 0.015526303090155125, + 0.012029229663312435, + 0.014230174012482166, + 0.07347651571035385, + 0.017494944855570793, + 0.020153755322098732, + 0.016185961663722992, + 0.019382275640964508, + 0.16979438066482544, + 0.016400713473558426, + 0.13624544441699982, + 0.0032434025779366493, + 0.003973393235355616, + 0.0025164103135466576, + 0.0027948960196226835, + 0.004005188122391701, + 0.0034311944618821144, + 0.0020539257675409317, + 0.0033974326215684414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.02328725904226303 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002965196967125 + ], + "opposite_logprob_tokens": [ + -2.3502964973449707 + ], + "logprob_reward": -0.1002965196967125, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002965196967125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023205114528536797, + 0.2166474312543869, + 0.03198022022843361, + 0.08953264355659485, + 0.14671866595745087, + 0.01717800460755825, + 0.033501144498586655, + 0.09572722017765045, + 0.2785751521587372, + 0.06693438440561295 + ], + "credit_map": [ + 0.003662681207060814, + 0.002757495967671275, + 0.003192613134160638, + 0.002579344902187586, + 0.002687733620405197, + 0.0027193555142730474, + 0.0030176998116075993, + 0.0025881933979690075, + 0.01923973858356476, + 0.010278528556227684, + 0.005242043640464544, + 0.0055496240966022015, + 0.0874980017542839, + 0.06344873458147049, + 0.013321258127689362, + 0.012069501914083958, + 0.0036048809997737408, + 0.0019262516871094704, + 0.00508189806714654, + 0.0028157271444797516, + 0.004232251085340977, + 0.007934359833598137, + 0.002091738861054182, + 0.004293116275221109, + 0.009793668054044247, + 0.010000585578382015, + 0.013064108788967133, + 0.007598823867738247, + 0.007787622511386871, + 0.014888797886669636, + 0.007856793701648712, + 0.018542245030403137, + 0.008228485472500324, + 0.006460940465331078, + 0.003283068770542741, + 0.004644304513931274, + 0.008006294257938862, + 0.05218225717544556, + 0.005980862770229578, + 0.05793246254324913, + 0.002178131602704525, + 0.0017680771416053176, + 0.0024723298847675323, + 0.0013079502386972308, + 0.0018706029513850808, + 0.0027933185920119286, + 0.002286083996295929, + 0.0025015121791511774, + 0.004200550261884928, + 0.005220992024987936, + 0.003185521811246872, + 0.002969004213809967, + 0.005272316746413708, + 0.0037556025199592113, + 0.005260836333036423, + 0.003636324545368552, + 0.011249042116105556, + 0.014157216064631939, + 0.008911005221307278, + 0.008712776936590672, + 0.013810336589813232, + 0.014662286266684532, + 0.011627055704593658, + 0.012597505934536457, + 0.04221993312239647, + 0.010598673485219479, + 0.012625755742192268, + 0.012542801909148693, + 0.012050656601786613, + 0.09860948473215103, + 0.010555894114077091, + 0.07937194406986237, + 0.0125083914026618, + 0.010653802193701267, + 0.006323462817817926, + 0.004447042942047119, + 0.0052300021052360535, + 0.014451445080339909, + 0.0028752526268363, + 0.01044499408453703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.1002965196967125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a dense field of slender, elongated fountain bamboo canes bracing themselves against a powerful gale. The entire scene is rendered in stunning ultra-high-definition slow motion, imbuing the natural event with the dramatic intensity of a high-speed sports documentary. Beneath a turbulent, dark-gray sky, each vibrantly green, segmented cane, its waxy surface gleaming, bends in a deep, tense arc, nearly driven to the ground by an unseen, overwhelming force. The camera maintains focus as the canes reach their elastic limit, drawing back with a violent, elastic recoil, their narrow leaves fluttering chaotically in a deep green blur. Dramatic, high-contrast light occasionally penetrates the fast-moving clouds, casting fleeting, sharp highlights on the tense stalks and individual water droplets flung from the leaves, freezing the moment of impact and release. The atmosphere is one of a fierce, primal struggle, showcasing the potent contest between the wind's brute force and the bamboo's profound flexibility.", + "seed": 28839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28839_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": -0.039626027923077345, + "vlm_logprob": -0.039626027923077345, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.039626027923077345, + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0068361214362084866 + ], + "opposite_logprob_tokens": [ + -6.1318359375 + ], + "logprob_reward": -0.0068361214362084866, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0068361214362084866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38135993480682373, + 0.0620310977101326, + 0.1197446957230568, + 0.04242943599820137, + 0.12420487403869629, + 0.0465572215616703, + 0.07513480633497238, + 0.0568731389939785, + 0.04834790900349617, + 0.04331693425774574 + ], + "credit_map": [ + 0.04675975814461708, + 0.04244363307952881, + 0.055880896747112274, + 0.036841146647930145, + 0.06729386001825333, + 0.03669597953557968, + 0.06444260478019714, + 0.03100203163921833, + 0.007900748401880264, + 0.008021572604775429, + 0.008184673264622688, + 0.006268445402383804, + 0.011354520916938782, + 0.008416393771767616, + 0.005134118255227804, + 0.0067506227642297745, + 0.015135414898395538, + 0.014111158438026905, + 0.013184662908315659, + 0.012038761749863625, + 0.011538779363036156, + 0.031158803030848503, + 0.008670443668961525, + 0.01390666700899601, + 0.005969394464045763, + 0.005379694979637861, + 0.005647661630064249, + 0.004234299063682556, + 0.007715849671512842, + 0.005998312029987574, + 0.0026587990578264, + 0.004825427662581205, + 0.030130930244922638, + 0.022519616410136223, + 0.008543813601136208, + 0.004609114024788141, + 0.03887699544429779, + 0.009027346968650818, + 0.0034775743260979652, + 0.00701948581263423, + 0.009035792201757431, + 0.006669935304671526, + 0.006128233857452869, + 0.004245517775416374, + 0.010475369170308113, + 0.004358228761702776, + 0.0022530267015099525, + 0.0033911147620528936, + 0.008173956535756588, + 0.01245841197669506, + 0.009026452898979187, + 0.0060808551497757435, + 0.01997586153447628, + 0.005844980012625456, + 0.0073201293125748634, + 0.006254155188798904, + 0.006369192618876696, + 0.0057368031702935696, + 0.008992316201329231, + 0.006469253916293383, + 0.008717208169400692, + 0.009450220502912998, + 0.0065307323820889, + 0.004607411101460457, + 0.006730956956744194, + 0.0035641861613839865, + 0.011475327424705029, + 0.0064275190234184265, + 0.005005175713449717, + 0.006658651866018772, + 0.0033412401098757982, + 0.00514485128223896, + 0.006571170408278704, + 0.004541154485195875, + 0.005018272437155247, + 0.004649149253964424, + 0.006920135114341974, + 0.00575686153024435, + 0.004314964171499014, + 0.005545224528759718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0068361214362084866 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307493194937706 + ], + "opposite_logprob_tokens": [ + -3.1680748462677 + ], + "logprob_reward": -0.04307493194937706, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307493194937706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06521552056074142, + 0.10496159642934799, + 0.13399596512317657, + 0.051044005900621414, + 0.44525036215782166, + 0.06578496843576431, + 0.029771676287055016, + 0.04752819240093231, + 0.04162382334470749, + 0.014823839999735355 + ], + "credit_map": [ + 0.008357817307114601, + 0.007152842823415995, + 0.010490830056369305, + 0.0068733482621610165, + 0.009061108343303204, + 0.00807433482259512, + 0.008242249488830566, + 0.006962982937693596, + 0.01360563188791275, + 0.01166390161961317, + 0.014583313837647438, + 0.009770466014742851, + 0.011680861935019493, + 0.02365882322192192, + 0.008093759417533875, + 0.011904830113053322, + 0.017815107479691505, + 0.011972770094871521, + 0.020230336114764214, + 0.012663641013205051, + 0.012507623061537743, + 0.02969597838819027, + 0.01015977282077074, + 0.018950743600726128, + 0.00759374164044857, + 0.005668430123478174, + 0.00882409606128931, + 0.005986141972243786, + 0.005012822803109884, + 0.005381491966545582, + 0.003051241859793663, + 0.00952603854238987, + 0.10652218014001846, + 0.09191696345806122, + 0.027541013434529305, + 0.01169294398277998, + 0.15311487019062042, + 0.02320217527449131, + 0.008804451674222946, + 0.022455796599388123, + 0.020595066249370575, + 0.021418211981654167, + 0.004443680867552757, + 0.002294242847710848, + 0.00910500530153513, + 0.0034118436742573977, + 0.0013491023564711213, + 0.003167810384184122, + 0.003905713092535734, + 0.0034787626937031746, + 0.003575935261324048, + 0.0029540532268583775, + 0.005574590992182493, + 0.0036012979689985514, + 0.0024674367159605026, + 0.004213887266814709, + 0.005720045883208513, + 0.00524302339181304, + 0.008926465176045895, + 0.004048184026032686, + 0.008182307705283165, + 0.006807642988860607, + 0.00311879045329988, + 0.0054817297495901585, + 0.0054474794305861, + 0.0029796536546200514, + 0.01016711164265871, + 0.005153694190084934, + 0.004426822531968355, + 0.004540408030152321, + 0.004572301171720028, + 0.004336354788392782, + 0.0017624985193833709, + 0.0013766322517767549, + 0.001972983358427882, + 0.002053032862022519, + 0.0020507676526904106, + 0.0022068340331315994, + 0.0015669706044718623, + 0.0018341209506615996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.04307493194937706 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288656026124954 + ], + "opposite_logprob_tokens": [ + -3.7732887268066406 + ], + "logprob_reward": -0.023288656026124954, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288656026124954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049420926719903946, + 0.16276101768016815, + 0.055953577160835266, + 0.03386692702770233, + 0.47531968355178833, + 0.10579383373260498, + 0.015416741371154785, + 0.02097863331437111, + 0.024365464225411415, + 0.05612323805689812 + ], + "credit_map": [ + 0.0060021523386240005, + 0.004778744652867317, + 0.006420725490897894, + 0.005160508211702108, + 0.007811905350536108, + 0.008756214752793312, + 0.005153088364750147, + 0.005337587092071772, + 0.019040606915950775, + 0.01634054258465767, + 0.023266546428203583, + 0.013893907889723778, + 0.01816920004785061, + 0.04054071754217148, + 0.017555730417370796, + 0.01395376492291689, + 0.007073563057929277, + 0.00529294041916728, + 0.0075629535131156445, + 0.005448020529001951, + 0.005747826304286718, + 0.012749144807457924, + 0.004553038161247969, + 0.007526090368628502, + 0.004648705944418907, + 0.004492964129894972, + 0.004148488864302635, + 0.0034058112651109695, + 0.006386320572346449, + 0.004654148127883673, + 0.0021859605330973864, + 0.003944525495171547, + 0.11634890735149384, + 0.09428553283214569, + 0.031043333932757378, + 0.015519115142524242, + 0.16052983701229095, + 0.025066865608096123, + 0.009260304272174835, + 0.02326580137014389, + 0.01141725480556488, + 0.03337199240922928, + 0.007623937912285328, + 0.009284308180212975, + 0.026684632524847984, + 0.007037574891000986, + 0.00403961818665266, + 0.0063345180824398994, + 0.0018785119755193591, + 0.0014460626989603043, + 0.001731502590700984, + 0.001992262667044997, + 0.002385402098298073, + 0.001956936437636614, + 0.0016101726796478033, + 0.002415890572592616, + 0.0028393517713993788, + 0.0028421024326235056, + 0.002830883488059044, + 0.0030493217054754496, + 0.004323052708059549, + 0.0016039591282606125, + 0.0014215538976714015, + 0.0020684076007455587, + 0.0027691901195794344, + 0.0026866672560572624, + 0.0032747574150562286, + 0.0024775369092822075, + 0.004542721901088953, + 0.0030551389791071415, + 0.0030290682334452868, + 0.0025303817819803953, + 0.007546448614448309, + 0.004208891652524471, + 0.009528795257210732, + 0.008233870379626751, + 0.005661880597472191, + 0.007043418474495411, + 0.0055449288338422775, + 0.008355006575584412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.023288656026124954 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054939065128564835 + ], + "opposite_logprob_tokens": [ + -2.929939031600952 + ], + "logprob_reward": -0.054939065128564835, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054939065128564835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10848403722047806, + 0.05414436385035515, + 0.11522742360830307, + 0.1012459546327591, + 0.1617039293050766, + 0.35777705907821655, + 0.017133139073848724, + 0.036769479513168335, + 0.019225120544433594, + 0.0282894317060709 + ], + "credit_map": [ + 0.012355716899037361, + 0.013158881105482578, + 0.017436139285564423, + 0.008316987194120884, + 0.01733606494963169, + 0.018228061497211456, + 0.01220190990716219, + 0.009450290352106094, + 0.0070406002923846245, + 0.005650840234011412, + 0.007407123222947121, + 0.005137435160577297, + 0.005621001590043306, + 0.014044607989490032, + 0.00423550046980381, + 0.005007260013371706, + 0.016408104449510574, + 0.013434328138828278, + 0.01571735367178917, + 0.011088068597018719, + 0.016847936436533928, + 0.021597351878881454, + 0.00828197505325079, + 0.011852310039103031, + 0.014437125064432621, + 0.015741653740406036, + 0.010707084089517593, + 0.007397105917334557, + 0.024971792474389076, + 0.013499941676855087, + 0.006137009710073471, + 0.008354258723556995, + 0.04246598854660988, + 0.02516425959765911, + 0.020196732133626938, + 0.013672818429768085, + 0.028477510437369347, + 0.012676578015089035, + 0.006328390911221504, + 0.012721670791506767, + 0.06635432690382004, + 0.1371021866798401, + 0.014896134845912457, + 0.010331287980079651, + 0.09538869559764862, + 0.014710726216435432, + 0.0068640015088021755, + 0.012129726819694042, + 0.0022684328723698854, + 0.002266994910314679, + 0.002377605764195323, + 0.0016240390250459313, + 0.0024513318203389645, + 0.0025486547965556383, + 0.0016176250064745545, + 0.001978455577045679, + 0.004422908648848534, + 0.004751884378492832, + 0.006773078814148903, + 0.002931375754997134, + 0.008122834376990795, + 0.0032149357721209526, + 0.0020991950295865536, + 0.004453267436474562, + 0.0022800015285611153, + 0.0015460955910384655, + 0.004032795317471027, + 0.0021811428014189005, + 0.0023876838386058807, + 0.0021698207128793, + 0.002673826878890395, + 0.0019537564367055893, + 0.004562086891382933, + 0.0020737231243401766, + 0.005087327677756548, + 0.003199793631210923, + 0.002103610895574093, + 0.0037601510994136333, + 0.0029604698065668344, + 0.004542273003607988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.054939065128564835 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999136507511139 + ], + "opposite_logprob_tokens": [ + -2.69499135017395 + ], + "logprob_reward": -0.06999136507511139, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999136507511139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06042127683758736, + 0.10320371389389038, + 0.04758885130286217, + 0.031442299485206604, + 0.5275385975837708, + 0.12895503640174866, + 0.018658164888620377, + 0.033619996160268784, + 0.02059810981154442, + 0.027973968535661697 + ], + "credit_map": [ + 0.008809207007288933, + 0.005948028992861509, + 0.00844742264598608, + 0.0058993506245315075, + 0.00631286483258009, + 0.008489487692713737, + 0.008734597824513912, + 0.007780319545418024, + 0.011996684595942497, + 0.010546223260462284, + 0.014015980996191502, + 0.009074465371668339, + 0.010499941185116768, + 0.02722826413810253, + 0.010985640808939934, + 0.008856523782014847, + 0.006065641064196825, + 0.005990907084196806, + 0.00724654458463192, + 0.0037178127095103264, + 0.006242456845939159, + 0.009857864119112492, + 0.003580773714929819, + 0.0048868549056351185, + 0.004356634337455034, + 0.003232826478779316, + 0.006147819571197033, + 0.003380716312676668, + 0.0036743339151144028, + 0.003508252091705799, + 0.0026683767791837454, + 0.004473340231925249, + 0.11575868725776672, + 0.11893071979284286, + 0.03144756704568863, + 0.014263649471104145, + 0.18472126126289368, + 0.027188964188098907, + 0.010159933008253574, + 0.02506779320538044, + 0.024022160097956657, + 0.049979355186223984, + 0.005156451836228371, + 0.0051767537370324135, + 0.03355787694454193, + 0.00352743873372674, + 0.0028146649710834026, + 0.0047203414142131805, + 0.0019502151990309358, + 0.0029074877966195345, + 0.0016679575201123953, + 0.0016402762848883867, + 0.00528594059869647, + 0.0018514757975935936, + 0.0015681477962061763, + 0.0017866658745333552, + 0.004493220709264278, + 0.00388159672729671, + 0.005994707345962524, + 0.0028023626655340195, + 0.006640306673943996, + 0.0029379946645349264, + 0.003332470078021288, + 0.003537343116477132, + 0.0026063527911901474, + 0.0019334116950631142, + 0.0036519623827189207, + 0.0019055074080824852, + 0.002925511682406068, + 0.002291574142873287, + 0.0031906431540846825, + 0.0020931477192789316, + 0.003323798533529043, + 0.002634751843288541, + 0.003835547249764204, + 0.004224452655762434, + 0.00392079958692193, + 0.0030625518411397934, + 0.002554523292928934, + 0.004417542368173599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.06999136507511139 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a dense field of slender, elongated fountain bamboo canes bracing themselves against a powerful gale. The entire scene is rendered in stunning ultra-high-definition slow motion, imbuing the natural event with the dramatic intensity of a high-speed sports documentary. Beneath a turbulent, dark-gray sky, each vibrantly green, segmented cane, its waxy surface gleaming, bends in a deep, tense arc, nearly driven to the ground by an unseen, overwhelming force. The camera maintains focus as the canes reach their elastic limit, drawing back with a violent, elastic recoil, their narrow leaves fluttering chaotically in a deep green blur. Dramatic, high-contrast light occasionally penetrates the fast-moving clouds, casting fleeting, sharp highlights on the tense stalks and individual water droplets flung from the leaves, freezing the moment of impact and release. The atmosphere is one of a fierce, primal struggle, showcasing the potent contest between the wind's brute force and the bamboo's profound flexibility.", + "seed": 28840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28840_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": -0.021781889349222185, + "vlm_logprob": -0.021781889349222185, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.021781889349222185, + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011991907842457294 + ], + "opposite_logprob_tokens": [ + -5.01199197769165 + ], + "logprob_reward": -0.011991907842457294, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011991907842457294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1371825635433197, + 0.08272344619035721, + 0.05939752981066704, + 0.21059687435626984, + 0.03892501816153526, + 0.091995470225811, + 0.14544042944908142, + 0.145791158080101, + 0.05453834310173988, + 0.033409152179956436 + ], + "credit_map": [ + 0.02343827113509178, + 0.008490560576319695, + 0.0215994194149971, + 0.009938626550137997, + 0.017844583839178085, + 0.01939883828163147, + 0.014129537157714367, + 0.02234272100031376, + 0.008411560207605362, + 0.007414848543703556, + 0.007444996386766434, + 0.00897691398859024, + 0.02341637946665287, + 0.011153277941048145, + 0.009346991777420044, + 0.006558472756296396, + 0.004127419553697109, + 0.005471520591527224, + 0.0029298339504748583, + 0.003750883275642991, + 0.018796920776367188, + 0.01258858386427164, + 0.00544928852468729, + 0.006283080670982599, + 0.028373483568429947, + 0.010550448670983315, + 0.022930661216378212, + 0.015540829859673977, + 0.021225856617093086, + 0.033061254769563675, + 0.03585108369588852, + 0.04306326061487198, + 0.004261845722794533, + 0.0031763019505888224, + 0.0058210259303450584, + 0.005793460179120302, + 0.005000591278076172, + 0.005229459144175053, + 0.00471287127584219, + 0.004929461516439915, + 0.010203019715845585, + 0.009294272400438786, + 0.008698520250618458, + 0.006028313189744949, + 0.017585789784789085, + 0.01868434064090252, + 0.010774516500532627, + 0.010726702399551868, + 0.014810671098530293, + 0.016222791746258736, + 0.022619349882006645, + 0.023774202913045883, + 0.019028062000870705, + 0.01835300400853157, + 0.01832052320241928, + 0.012311824597418308, + 0.01783842220902443, + 0.011694329790771008, + 0.023600999265909195, + 0.013971016742289066, + 0.026308385655283928, + 0.018147559836506844, + 0.018476393073797226, + 0.015754055231809616, + 0.006211745552718639, + 0.0059926738031208515, + 0.006418315693736076, + 0.006139854900538921, + 0.0058806585147976875, + 0.008327480405569077, + 0.006685057654976845, + 0.008882559835910797, + 0.0047316839918494225, + 0.004159686621278524, + 0.003768304130062461, + 0.0022658868692815304, + 0.005984251853078604, + 0.004551352001726627, + 0.004229004494845867, + 0.0037189803551882505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.011991907842457294 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710893243551254 + ], + "opposite_logprob_tokens": [ + -3.4087109565734863 + ], + "logprob_reward": -0.033710893243551254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710893243551254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1645638644695282, + 0.07339254021644592, + 0.0970345214009285, + 0.13205920159816742, + 0.025019671767950058, + 0.11551498621702194, + 0.1186402440071106, + 0.15564106404781342, + 0.05757715925574303, + 0.06055672839283943 + ], + "credit_map": [ + 0.02338332124054432, + 0.010758204385638237, + 0.011901741847395897, + 0.008458280935883522, + 0.03619350865483284, + 0.04019910469651222, + 0.01080181822180748, + 0.02286788634955883, + 0.0063211312517523766, + 0.006012954283505678, + 0.005326870363205671, + 0.005986619740724564, + 0.019943827763199806, + 0.011182785965502262, + 0.01011018268764019, + 0.008508175611495972, + 0.005293895024806261, + 0.006357449572533369, + 0.004657011479139328, + 0.005110345780849457, + 0.02769090048968792, + 0.021510830149054527, + 0.009252081625163555, + 0.017162004485726357, + 0.014990568161010742, + 0.008127564564347267, + 0.013182300142943859, + 0.012729061767458916, + 0.020626189187169075, + 0.02754826471209526, + 0.014387036673724651, + 0.020468225702643394, + 0.002855052938684821, + 0.003137609688565135, + 0.002141964389011264, + 0.0021772587206214666, + 0.005002663470804691, + 0.004303611349314451, + 0.0021729636937379837, + 0.0032285479828715324, + 0.010241411626338959, + 0.013452883809804916, + 0.013392205350100994, + 0.008272760547697544, + 0.020247062668204308, + 0.02382473833858967, + 0.014178259298205376, + 0.0119056710973382, + 0.010430742055177689, + 0.012014254927635193, + 0.021480267867445946, + 0.017654890194535255, + 0.018898827955126762, + 0.015387196093797684, + 0.012818976305425167, + 0.00995510071516037, + 0.015002693049609661, + 0.01840279810130596, + 0.017933588474988937, + 0.026635820046067238, + 0.020129438489675522, + 0.01896093226969242, + 0.01851697824895382, + 0.02005881816148758, + 0.005321730859577656, + 0.00586744723841548, + 0.007420148700475693, + 0.005748762749135494, + 0.006892715115100145, + 0.009180812165141106, + 0.007028954103589058, + 0.01011659111827612, + 0.006952201947569847, + 0.004191302228718996, + 0.003999905660748482, + 0.0035672562662512064, + 0.006763465236872435, + 0.018382644280791283, + 0.004003580193966627, + 0.012696376070380211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.033710893243551254 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637686260044575 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637686260044575, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637686260044575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2723771333694458, + 0.05674663558602333, + 0.11804606765508652, + 0.12392885237932205, + 0.10873370617628098, + 0.069684237241745, + 0.056945886462926865, + 0.07322561740875244, + 0.02366841956973076, + 0.09664346277713776 + ], + "credit_map": [ + 0.04297294840216637, + 0.03288499638438225, + 0.025795338675379753, + 0.024827132001519203, + 0.038502227514982224, + 0.04641902074217796, + 0.02144225686788559, + 0.03953322023153305, + 0.006966488901525736, + 0.006453778129070997, + 0.0062930951826274395, + 0.006425121333450079, + 0.00955590233206749, + 0.0073320334777235985, + 0.006871635094285011, + 0.006848582532256842, + 0.010184754617512226, + 0.01294313557446003, + 0.008294042199850082, + 0.006786007899791002, + 0.04021565243601799, + 0.01627477817237377, + 0.01229556743055582, + 0.01105212327092886, + 0.013825732283294201, + 0.008080982603132725, + 0.008867917582392693, + 0.008004548959434032, + 0.018156329169869423, + 0.033327728509902954, + 0.014414804056286812, + 0.019250815734267235, + 0.01479229237884283, + 0.013861758634448051, + 0.00920588057488203, + 0.008555048145353794, + 0.022799460217356682, + 0.017956310883164406, + 0.010828708298504353, + 0.010734239593148232, + 0.007835048250854015, + 0.005954793654382229, + 0.006330783013254404, + 0.007093269377946854, + 0.014645924791693687, + 0.00868441816419363, + 0.009955371730029583, + 0.009184633381664753, + 0.004750876221805811, + 0.0049197012558579445, + 0.006688011810183525, + 0.0065362113527953625, + 0.008838405832648277, + 0.007887101732194424, + 0.008770848624408245, + 0.008554725907742977, + 0.009652867913246155, + 0.007264541462063789, + 0.009186000563204288, + 0.00917237438261509, + 0.01268282812088728, + 0.009239903651177883, + 0.008643707260489464, + 0.007383388001471758, + 0.0027087382040917873, + 0.0021468601189553738, + 0.003147569950670004, + 0.0027951146475970745, + 0.0025146976113319397, + 0.0030815675854682922, + 0.003183641703799367, + 0.004090229980647564, + 0.012174292467534542, + 0.010588927194476128, + 0.0073798769153654575, + 0.005678331013768911, + 0.011210194788873196, + 0.025822093710303307, + 0.008166627027094364, + 0.015623118728399277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008637686260044575 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943363532423973 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.005943363532423973, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943363532423973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12052086740732193, + 0.2677442729473114, + 0.07944313436746597, + 0.04760442674160004, + 0.07704703509807587, + 0.13172942399978638, + 0.08201972395181656, + 0.09126783162355423, + 0.02563636191189289, + 0.0769869014620781 + ], + "credit_map": [ + 0.02021896466612816, + 0.009371565654873848, + 0.011640455573797226, + 0.008962824940681458, + 0.017264610156416893, + 0.027379607781767845, + 0.0073328702710568905, + 0.01834997721016407, + 0.02675313875079155, + 0.021170493215322495, + 0.018836699426174164, + 0.028079060837626457, + 0.07846393436193466, + 0.03982919827103615, + 0.03234666585922241, + 0.02226509153842926, + 0.005618850234895945, + 0.0070327953435480595, + 0.005281846970319748, + 0.005046195350587368, + 0.017708130180835724, + 0.017538372427225113, + 0.006543139927089214, + 0.014673800207674503, + 0.004430919419974089, + 0.005052223335951567, + 0.004638184327632189, + 0.0057629127986729145, + 0.007910536602139473, + 0.006119647063314915, + 0.0077800266444683075, + 0.005909978877753019, + 0.00928223691880703, + 0.010771135799586773, + 0.005631429608911276, + 0.006123635917901993, + 0.01646805740892887, + 0.01261372584849596, + 0.007579598110169172, + 0.008577205240726471, + 0.015726201236248016, + 0.015079867094755173, + 0.0128653384745121, + 0.015090923756361008, + 0.016315890476107597, + 0.022766610607504845, + 0.01769901067018509, + 0.0161855760961771, + 0.009091194719076157, + 0.00749039463698864, + 0.00815805047750473, + 0.00789973046630621, + 0.014602246694266796, + 0.01201406866312027, + 0.009667188860476017, + 0.013096851296722889, + 0.009529627859592438, + 0.008762186393141747, + 0.014982429333031178, + 0.010000001639127731, + 0.016858957707881927, + 0.011740785092115402, + 0.010332814417779446, + 0.009061025455594063, + 0.002409168751910329, + 0.0032037533819675446, + 0.002088220091536641, + 0.0016135642072185874, + 0.00563711766153574, + 0.004479575902223587, + 0.0032222492154687643, + 0.0029827130492776632, + 0.008233570493757725, + 0.013720773160457611, + 0.008362219668924809, + 0.007972641848027706, + 0.01032521016895771, + 0.008220627903938293, + 0.009375846944749355, + 0.01077601220458746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.005943363532423973 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862559586763382 + ], + "opposite_logprob_tokens": [ + -3.0486257076263428 + ], + "logprob_reward": -0.04862559586763382, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862559586763382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1666143536567688, + 0.11130855977535248, + 0.0762811005115509, + 0.08399851620197296, + 0.08558131009340286, + 0.14412857592105865, + 0.11921887844800949, + 0.09562451392412186, + 0.022460656240582466, + 0.09478355944156647 + ], + "credit_map": [ + 0.019027676433324814, + 0.012217062525451183, + 0.017040690407156944, + 0.010097986087203026, + 0.026677213609218597, + 0.027929002419114113, + 0.01717868633568287, + 0.03644603490829468, + 0.013113421387970448, + 0.008839651010930538, + 0.00957725290209055, + 0.011170901358127594, + 0.027184689417481422, + 0.014837659895420074, + 0.015312010422348976, + 0.011272968724370003, + 0.005591448396444321, + 0.007383286952972412, + 0.005909896921366453, + 0.004142757970839739, + 0.009040748700499535, + 0.02459026128053665, + 0.004538672044873238, + 0.01508402731269598, + 0.006850884761661291, + 0.006725642830133438, + 0.003984932322055101, + 0.006055546458810568, + 0.014142551459372044, + 0.026682980358600616, + 0.007900978438556194, + 0.011654999107122421, + 0.00811154767870903, + 0.010015314444899559, + 0.005775278899818659, + 0.008219956420361996, + 0.01990838535130024, + 0.014374732971191406, + 0.010066461749374866, + 0.009109629318118095, + 0.015067042782902718, + 0.013942860998213291, + 0.016749782487750053, + 0.014798984862864017, + 0.02645927481353283, + 0.025213979184627533, + 0.01686384528875351, + 0.015032787807285786, + 0.0152013273909688, + 0.00930938683450222, + 0.020478051155805588, + 0.0115993432700634, + 0.019716061651706696, + 0.015548438765108585, + 0.012051057070493698, + 0.015315203927457333, + 0.010434412397444248, + 0.00979615468531847, + 0.011011997237801552, + 0.010359887033700943, + 0.01748056337237358, + 0.013640958815813065, + 0.013121756725013256, + 0.009778780862689018, + 0.0027505417820066214, + 0.0017505846917629242, + 0.003605707548558712, + 0.00199904409237206, + 0.0038696574047207832, + 0.003408261341974139, + 0.002274587517604232, + 0.002802271395921707, + 0.014017552137374878, + 0.014455939643085003, + 0.006908728741109371, + 0.007553811650723219, + 0.014890172518789768, + 0.012029302306473255, + 0.009209328331053257, + 0.015718718990683556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04862559586763382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static close-up captures a dense field of slender, elongated fountain bamboo canes bracing themselves against a powerful gale. The entire scene is rendered in stunning ultra-high-definition slow motion, imbuing the natural event with the dramatic intensity of a high-speed sports documentary. Beneath a turbulent, dark-gray sky, each vibrantly green, segmented cane, its waxy surface gleaming, bends in a deep, tense arc, nearly driven to the ground by an unseen, overwhelming force. The camera maintains focus as the canes reach their elastic limit, drawing back with a violent, elastic recoil, their narrow leaves fluttering chaotically in a deep green blur. Dramatic, high-contrast light occasionally penetrates the fast-moving clouds, casting fleeting, sharp highlights on the tense stalks and individual water droplets flung from the leaves, freezing the moment of impact and release. The atmosphere is one of a fierce, primal struggle, showcasing the potent contest between the wind's brute force and the bamboo's profound flexibility.", + "seed": 28841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_3/index_6239_seed_28841_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": -0.0499748298432678, + "vlm_logprob": -0.0499748298432678, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0499748298432678, + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19146060943603516 + ], + "opposite_logprob_tokens": [ + -1.8164606094360352 + ], + "logprob_reward": -0.19146060943603516, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19146060943603516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22058354318141937, + 0.10681939125061035, + 0.060160040855407715, + 0.027828393504023552, + 0.057435858994722366, + 0.08153865486383438, + 0.028124116361141205, + 0.030276257544755936, + 0.13847824931144714, + 0.24875551462173462 + ], + "credit_map": [ + 0.03159954398870468, + 0.04231281578540802, + 0.019900185987353325, + 0.01626899279654026, + 0.037319377064704895, + 0.026599079370498657, + 0.026132727041840553, + 0.020450828596949577, + 0.01741807535290718, + 0.03372231125831604, + 0.0032064684201031923, + 0.002652880735695362, + 0.03066820465028286, + 0.010997498407959938, + 0.0030807731673121452, + 0.0050731804221868515, + 0.00623016431927681, + 0.00855073519051075, + 0.004660764709115028, + 0.007243604864925146, + 0.014144339598715305, + 0.009765712544322014, + 0.004737719427794218, + 0.004827000666409731, + 0.002493883715942502, + 0.003282683901488781, + 0.00252953520976007, + 0.0017912612529471517, + 0.006162622012197971, + 0.003917363006621599, + 0.003418921958655119, + 0.004232123959809542, + 0.0063680666498839855, + 0.008692813105881214, + 0.005832234863191843, + 0.0050633675418794155, + 0.008294947445392609, + 0.009213910438120365, + 0.006992105394601822, + 0.006978412624448538, + 0.006692834198474884, + 0.004796831402927637, + 0.009572602808475494, + 0.006613609381020069, + 0.01013475563377142, + 0.01210104115307331, + 0.018044548109173775, + 0.013582431711256504, + 0.0029822122305631638, + 0.0031907716765999794, + 0.001981679815798998, + 0.0017435936024412513, + 0.003046533791348338, + 0.003452513599768281, + 0.006019706837832928, + 0.005707104690372944, + 0.002891340060159564, + 0.004040069878101349, + 0.0024272568989545107, + 0.002103891223669052, + 0.00409103836864233, + 0.0048862025141716, + 0.004741310141980648, + 0.005095148459076881, + 0.01863696798682213, + 0.014303961768746376, + 0.009255783632397652, + 0.007834360934793949, + 0.009356079623103142, + 0.021801404654979706, + 0.02092793583869934, + 0.03636174649000168, + 0.00951484963297844, + 0.005420053377747536, + 0.006001020781695843, + 0.005573569796979427, + 0.012317977845668793, + 0.10877101123332977, + 0.015551441349089146, + 0.08560557663440704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.19146060943603516 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331509441137314 + ], + "opposite_logprob_tokens": [ + -3.773315191268921 + ], + "logprob_reward": -0.02331509441137314, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331509441137314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23625372350215912, + 0.17618799209594727, + 0.04373495280742645, + 0.10339440405368805, + 0.06238012760877609, + 0.036603059619665146, + 0.028062010183930397, + 0.03748215362429619, + 0.09657811373472214, + 0.17932353913784027 + ], + "credit_map": [ + 0.04672427102923393, + 0.08372421562671661, + 0.007905114442110062, + 0.007558736018836498, + 0.055529847741127014, + 0.019301703199744225, + 0.007033095695078373, + 0.008476724848151207, + 0.025809308513998985, + 0.052572064101696014, + 0.007698438595980406, + 0.005681865848600864, + 0.048343248665332794, + 0.01780090294778347, + 0.0079331761226058, + 0.01034897193312645, + 0.004051088821142912, + 0.0052799805998802185, + 0.004609971772879362, + 0.0038425703532993793, + 0.006133731920272112, + 0.00762670673429966, + 0.003423688467592001, + 0.008767211809754372, + 0.0076478649862110615, + 0.013674750924110413, + 0.006551749538630247, + 0.006796833593398333, + 0.02671322040259838, + 0.015581863932311535, + 0.015745120123028755, + 0.010682989843189716, + 0.006076040677726269, + 0.007845094427466393, + 0.007353243418037891, + 0.006482163909822702, + 0.010793141089379787, + 0.010124598629772663, + 0.005985358264297247, + 0.0077204834669828415, + 0.003383512143045664, + 0.004359799437224865, + 0.0038062441162765026, + 0.003347401972860098, + 0.006189510691910982, + 0.004849622491747141, + 0.0053136213682591915, + 0.005353344138711691, + 0.0022875976283103228, + 0.0033755924087017775, + 0.0030800269450992346, + 0.002684107981622219, + 0.0033865098375827074, + 0.004102980252355337, + 0.004751632455736399, + 0.004393559880554676, + 0.0030262647196650505, + 0.00334860198199749, + 0.0029337776359170675, + 0.004320766776800156, + 0.003751967567950487, + 0.0049690864980220795, + 0.00949203409254551, + 0.005639650858938694, + 0.01244578417390585, + 0.0107562942430377, + 0.007481646258383989, + 0.0075692846439778805, + 0.006525822449475527, + 0.01761380210518837, + 0.013757442124187946, + 0.020428022369742393, + 0.0072228917852044106, + 0.0057912482880055904, + 0.00576976640149951, + 0.007056611590087414, + 0.00917875673621893, + 0.07599771022796631, + 0.011191200464963913, + 0.0571153424680233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02331509441137314 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821395406499505 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821395406499505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821395406499505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24162514507770538, + 0.15759101510047913, + 0.0899694636464119, + 0.0588960200548172, + 0.10007227957248688, + 0.07873280346393585, + 0.05647319182753563, + 0.051818396896123886, + 0.09297192096710205, + 0.07184968143701553 + ], + "credit_map": [ + 0.03047669678926468, + 0.028919002041220665, + 0.015994569286704063, + 0.022042755037546158, + 0.05363349989056587, + 0.03261587396264076, + 0.02750328741967678, + 0.030439475551247597, + 0.016012702137231827, + 0.030425596982240677, + 0.017325595021247864, + 0.01069519855082035, + 0.036312542855739594, + 0.019155165180563927, + 0.012612471356987953, + 0.015051760710775852, + 0.00984597485512495, + 0.011118533089756966, + 0.009927518665790558, + 0.009670985862612724, + 0.014524389989674091, + 0.01545425970107317, + 0.009057717397809029, + 0.010370089672505856, + 0.00615099398419261, + 0.00781095540151, + 0.0071744611486792564, + 0.004781651776283979, + 0.009053315035998821, + 0.008721784688532352, + 0.006879844237118959, + 0.008323016576468945, + 0.00788431242108345, + 0.011474407278001308, + 0.010480770841240883, + 0.008527922444045544, + 0.014428992755711079, + 0.01711520552635193, + 0.011613489128649235, + 0.018547184765338898, + 0.005415632389485836, + 0.006591652985662222, + 0.011083093471825123, + 0.0069695133715868, + 0.008527149446308613, + 0.015052303671836853, + 0.015066354535520077, + 0.010027114301919937, + 0.006066583096981049, + 0.0064658853225409985, + 0.00407885666936636, + 0.0048876116052269936, + 0.007271901238709688, + 0.00923867803066969, + 0.007343421224504709, + 0.011120260693132877, + 0.004172049462795258, + 0.0037811431102454662, + 0.005357845686376095, + 0.006106867454946041, + 0.00601781252771616, + 0.00905519723892212, + 0.007899080403149128, + 0.009428405202925205, + 0.013120140880346298, + 0.009597214870154858, + 0.0061760894022881985, + 0.005318682175129652, + 0.005834376905113459, + 0.015751875936985016, + 0.012759335339069366, + 0.024414213374257088, + 0.004710240755230188, + 0.006586492992937565, + 0.0043660299852490425, + 0.005714108236134052, + 0.006606635171920061, + 0.02026110142469406, + 0.005710020195692778, + 0.017895061522722244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002821395406499505 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944904405623674 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944904405623674, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944904405623674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13559386134147644, + 0.15555521845817566, + 0.06705430150032043, + 0.17502640187740326, + 0.03267746791243553, + 0.03301501274108887, + 0.03103245235979557, + 0.05275433138012886, + 0.14099079370498657, + 0.17630018293857574 + ], + "credit_map": [ + 0.02331937849521637, + 0.038352906703948975, + 0.011185658164322376, + 0.00843335036188364, + 0.021065296605229378, + 0.011286841705441475, + 0.012238148599863052, + 0.009712282568216324, + 0.0226390790194273, + 0.04538407176733017, + 0.007928336039185524, + 0.00497118104249239, + 0.04252288118004799, + 0.017472553998231888, + 0.005952011793851852, + 0.008685117587447166, + 0.007127268239855766, + 0.010992135852575302, + 0.008182654157280922, + 0.005936043336987495, + 0.008691479451954365, + 0.00782736111432314, + 0.00622352072969079, + 0.012073840014636517, + 0.012008769437670708, + 0.02245527133345604, + 0.013888600282371044, + 0.010608980432152748, + 0.03816717490553856, + 0.024623000994324684, + 0.02848655730485916, + 0.02478804625570774, + 0.0025791856460273266, + 0.00405459338799119, + 0.002790737897157669, + 0.003557646181434393, + 0.006375962868332863, + 0.0060872482135891914, + 0.003250586800277233, + 0.003981504589319229, + 0.004055745899677277, + 0.004339948296546936, + 0.0038571541663259268, + 0.002895879093557596, + 0.0053560477681458, + 0.004562010522931814, + 0.0037089702673256397, + 0.004239257425069809, + 0.0032945312559604645, + 0.003723748028278351, + 0.003204706823453307, + 0.0029047327116131783, + 0.0035198184195905924, + 0.003891996806487441, + 0.004552892874926329, + 0.005940027069300413, + 0.0032289321534335613, + 0.004035710357129574, + 0.0037468024529516697, + 0.004982460290193558, + 0.005697224289178848, + 0.00917072407901287, + 0.012125702574849129, + 0.009766774252057076, + 0.012240207754075527, + 0.014179840683937073, + 0.012968341819941998, + 0.012551918625831604, + 0.012440784834325314, + 0.029497915878891945, + 0.024406753480434418, + 0.02270502969622612, + 0.007134601008147001, + 0.005510126240551472, + 0.004448033403605223, + 0.004566491581499577, + 0.009482035413384438, + 0.07484842836856842, + 0.010834919288754463, + 0.05947554484009743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.005944904405623674 + }, + { + "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 low-angle shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera framing a static close-up of slender, elongated fountain bamboo canes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the bamboo canes vibrantly green with a waxy surface that gleams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the canes bending in a deep arc against a turbulent, dark-gray sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are water droplets visible flung from the leaves with sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332145556807518 + ], + "opposite_logprob_tokens": [ + -3.651332139968872 + ], + "logprob_reward": -0.026332145556807518, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332145556807518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.189909890294075, + 0.4051309823989868, + 0.047595951706171036, + 0.03839486092329025, + 0.03490309789776802, + 0.03748166561126709, + 0.019980447366833687, + 0.026611292734742165, + 0.1477813422679901, + 0.052210453897714615 + ], + "credit_map": [ + 0.03128982335329056, + 0.048152465373277664, + 0.010458637960255146, + 0.00995578896254301, + 0.03973350301384926, + 0.02398437261581421, + 0.013382735662162304, + 0.012952566146850586, + 0.040923938155174255, + 0.07368579506874084, + 0.030978331342339516, + 0.019695933908224106, + 0.12691804766654968, + 0.065041184425354, + 0.023303717374801636, + 0.024584032595157623, + 0.004293935839086771, + 0.005575032904744148, + 0.006106636952608824, + 0.006053199525922537, + 0.005656595341861248, + 0.008508265018463135, + 0.004391024820506573, + 0.007011259440332651, + 0.00326507817953825, + 0.005258596036583185, + 0.0038683649618178606, + 0.0027916308026760817, + 0.008579089306294918, + 0.005324006546288729, + 0.00466838339343667, + 0.004639711230993271, + 0.0030214725993573666, + 0.00464915856719017, + 0.004044214263558388, + 0.003813087474554777, + 0.006229157093912363, + 0.0045911031775176525, + 0.004428790882229805, + 0.004126113373786211, + 0.0023407936096191406, + 0.003343272488564253, + 0.004548362921923399, + 0.0038346825167536736, + 0.005322455428540707, + 0.0063240863382816315, + 0.006565273739397526, + 0.00520273856818676, + 0.0025342500302940607, + 0.002374674892053008, + 0.001482520718127489, + 0.0012914288090541959, + 0.0017734476132318377, + 0.0031206468120217323, + 0.0029068856965750456, + 0.004496593028306961, + 0.0024908671621233225, + 0.0023508877493441105, + 0.0021009547635912895, + 0.002546481555327773, + 0.0035024669487029314, + 0.003541252575814724, + 0.0054685454815626144, + 0.004609833937138319, + 0.020731372758746147, + 0.014679757878184319, + 0.011865274049341679, + 0.009245285764336586, + 0.008084078319370747, + 0.025600779801607132, + 0.019816458225250244, + 0.03775833919644356, + 0.004303661175072193, + 0.007759213447570801, + 0.004485758952796459, + 0.006938530132174492, + 0.00955275446176529, + 0.007408719975501299, + 0.006102789659053087, + 0.005659028422087431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.026332145556807518 + } + ], + "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/0000060/group_4/index_tensor([835, 835])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_tensor([835, 835])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aea75ce8fb09ea5f5552d140670142d4df4185cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_tensor([835, 835])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, captured from behind and slightly above, captures a man standing on the balcony of a modern high-rise apartment building. He wears a soft, dark gray cotton hoodie and leans slightly forward, his hands resting naturally on the cool black metal railing. His short hair ruffles gently in the evening breeze, and our gaze drifts over his shoulder, focused on his intense gaze. Below, a busy street buzzes with traffic. At dusk, countless headlights and taillights converge into flowing rivers of light, while miniature pedestrians hurried across the sidewalk. The last rays of sunset paint the sky with warm orange-purple hues, contrasting sharply with the gradually brightening cool lights of the city buildings. The complex light highlights the stern profile of his face and the fine textures of his clothing. The entire image has a photographic quality, and the strong depth of field isolates his quiet contemplation from the bustling city below, creating an atmosphere of urban contemplation that is both detached and profound.", + "seed": 23648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_835_seed_23648_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": -0.027464228169992565, + "vlm_logprob": -0.027464228169992565, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.027464228169992565, + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.02202339470386505 + ], + "opposite_logprob_tokens": [ + -4.1470232009887695 + ], + "logprob_reward": -0.02202339470386505, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02202339470386505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4024616777896881, + 0.06972815841436386, + 0.044336944818496704, + 0.06081356853246689, + 0.04830954968929291, + 0.058220017701387405, + 0.14765377342700958, + 0.02329307794570923, + 0.11003980785608292, + 0.03514339402318001 + ], + "credit_map": [ + 0.03933781012892723, + 0.035388022661209106, + 0.04765124246478081, + 0.058897893875837326, + 0.047669000923633575, + 0.06467597931623459, + 0.03743678703904152, + 0.07140496373176575, + 0.011023543775081635, + 0.007779828738421202, + 0.009695273824036121, + 0.011083623394370079, + 0.007445028517395258, + 0.008937543258070946, + 0.004957951605319977, + 0.008805363439023495, + 0.005076482426375151, + 0.006210925057530403, + 0.0035297221038490534, + 0.006649733986705542, + 0.005021607968956232, + 0.00531799579039216, + 0.007483392022550106, + 0.00504708755761385, + 0.0064805494621396065, + 0.00635709660127759, + 0.006143949460238218, + 0.010111669078469276, + 0.0056441607885062695, + 0.008912317454814911, + 0.005913279484957457, + 0.011250547133386135, + 0.006687880493700504, + 0.00844352226704359, + 0.004323374945670366, + 0.006469293963164091, + 0.004464755300432444, + 0.007292237598448992, + 0.00496998755261302, + 0.005658499430865049, + 0.007068682461977005, + 0.005218797363340855, + 0.007973844185471535, + 0.008464435115456581, + 0.008239871822297573, + 0.00855527725070715, + 0.005077222362160683, + 0.007621884811669588, + 0.022882958874106407, + 0.01615193672478199, + 0.01837123930454254, + 0.02142123505473137, + 0.014822029508650303, + 0.01955701969563961, + 0.012959571555256844, + 0.02148779109120369, + 0.004051588475704193, + 0.002531081438064575, + 0.00256487843580544, + 0.0037843515165150166, + 0.0024877453688532114, + 0.002358867321163416, + 0.002670876681804657, + 0.0028436901047825813, + 0.014952491968870163, + 0.013084120117127895, + 0.01271173544228077, + 0.018295561894774437, + 0.013215804472565651, + 0.016537537798285484, + 0.009137955494225025, + 0.012104595080018044, + 0.0029689448419958353, + 0.005904879886657, + 0.0028606632258743048, + 0.005951750557869673, + 0.002853187033906579, + 0.0033662826754152775, + 0.004564886447042227, + 0.0066728005185723305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.02202339470386505 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.04307927191257477 + ], + "opposite_logprob_tokens": [ + -3.168079376220703 + ], + "logprob_reward": -0.04307927191257477, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307927191257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33924537897109985, + 0.1073954626917839, + 0.07631008327007294, + 0.08429382741451263, + 0.09545323997735977, + 0.06470269709825516, + 0.057937417179346085, + 0.058863330632448196, + 0.04589305818080902, + 0.06990554183721542 + ], + "credit_map": [ + 0.03917120024561882, + 0.06943236291408539, + 0.03140809014439583, + 0.06246485933661461, + 0.03260834515094757, + 0.03611622005701065, + 0.037120312452316284, + 0.03092399425804615, + 0.016904598101973534, + 0.017539719119668007, + 0.0116725480183959, + 0.01766282320022583, + 0.009225787594914436, + 0.012862796895205975, + 0.012221242301166058, + 0.009305951185524464, + 0.01299126073718071, + 0.010407673195004463, + 0.00850305799394846, + 0.013603865168988705, + 0.0075334059074521065, + 0.009256493300199509, + 0.007069202605634928, + 0.006945119239389896, + 0.013062835671007633, + 0.007282235659658909, + 0.011042848229408264, + 0.013364850543439388, + 0.007116606459021568, + 0.011176102794706821, + 0.009014726616442204, + 0.01223362423479557, + 0.014511710964143276, + 0.011747407726943493, + 0.01101035624742508, + 0.016391417011618614, + 0.007605313789099455, + 0.009317823685705662, + 0.01408038754016161, + 0.010788827203214169, + 0.011084903962910175, + 0.008188941515982151, + 0.0066850557923316956, + 0.0097450390458107, + 0.005799873732030392, + 0.008473040536046028, + 0.005908132530748844, + 0.008817706257104874, + 0.007120386231690645, + 0.009221946820616722, + 0.005793054588139057, + 0.00993660744279623, + 0.007788521703332663, + 0.006318800617009401, + 0.0052560074254870415, + 0.0065020970068871975, + 0.008953001350164413, + 0.009153653867542744, + 0.005893652327358723, + 0.00869719497859478, + 0.004355865065008402, + 0.005930330138653517, + 0.007209410425275564, + 0.008670218288898468, + 0.004022669047117233, + 0.004138939082622528, + 0.0032759797759354115, + 0.008282500319182873, + 0.004615291021764278, + 0.005646366160362959, + 0.005145246163010597, + 0.010766069404780865, + 0.006244062911719084, + 0.009461469948291779, + 0.006865456234663725, + 0.011216342449188232, + 0.0045894174836575985, + 0.007546560373157263, + 0.00910208374261856, + 0.014880145899951458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04307927191257477 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3957265615463257, + 0.07879186421632767, + 0.07105568051338196, + 0.10677539557218552, + 0.10613685846328735, + 0.05470690876245499, + 0.0615905299782753, + 0.05261467769742012, + 0.04093519225716591, + 0.03166627883911133 + ], + "credit_map": [ + 0.055794607847929, + 0.03703086078166962, + 0.041864391416311264, + 0.07761389017105103, + 0.04027538746595383, + 0.04518330469727516, + 0.03179186210036278, + 0.06617224961519241, + 0.01012041699141264, + 0.008066968061029911, + 0.00774286687374115, + 0.01396056916564703, + 0.008054531179368496, + 0.009689582511782646, + 0.007154850289225578, + 0.01400208193808794, + 0.010060238651931286, + 0.007824713364243507, + 0.008249432779848576, + 0.012540355324745178, + 0.009751897305250168, + 0.00782093033194542, + 0.005952851381152868, + 0.008855265565216541, + 0.016106311231851578, + 0.009361740201711655, + 0.01287367008626461, + 0.016599493101239204, + 0.01262049563229084, + 0.01174856722354889, + 0.012078709900379181, + 0.015386400744318962, + 0.01647176407277584, + 0.012094492092728615, + 0.013876920565962791, + 0.016405202448368073, + 0.009699076414108276, + 0.011908301152288914, + 0.010687966831028461, + 0.014993133023381233, + 0.005661915522068739, + 0.005872899200767279, + 0.005239262245595455, + 0.010858600959181786, + 0.004506845958530903, + 0.007132070604711771, + 0.0059966337867081165, + 0.009438679553568363, + 0.00855832826346159, + 0.010907965712249279, + 0.004804997239261866, + 0.009291746653616428, + 0.004981283564120531, + 0.008552909828722477, + 0.005839712917804718, + 0.00865358579903841, + 0.008116732351481915, + 0.008904394693672657, + 0.0055623468942940235, + 0.008344915695488453, + 0.004544015508145094, + 0.005514667369425297, + 0.005046091508120298, + 0.006581513676792383, + 0.004644612781703472, + 0.005359889939427376, + 0.0032334134448319674, + 0.006453776266425848, + 0.0031707314774394035, + 0.0058203511871397495, + 0.00472076004371047, + 0.007531654089689255, + 0.00429620873183012, + 0.003744056448340416, + 0.0021149080712348223, + 0.00544995442032814, + 0.0027172162663191557, + 0.0032276331912726164, + 0.004407892003655434, + 0.005708408076316118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0028174726758152246 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.04305654764175415 + ], + "opposite_logprob_tokens": [ + -3.1680564880371094 + ], + "logprob_reward": -0.04305654764175415, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305654764175415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41329172253608704, + 0.08476923406124115, + 0.05172897130250931, + 0.1289113312959671, + 0.07723823934793472, + 0.048882097005844116, + 0.03255699574947357, + 0.021167093887925148, + 0.03294627368450165, + 0.10850808024406433 + ], + "credit_map": [ + 0.04422983527183533, + 0.09445179998874664, + 0.03649207577109337, + 0.08699037879705429, + 0.030468761920928955, + 0.04204409196972847, + 0.03612774237990379, + 0.042486999183893204, + 0.011326401494443417, + 0.014049215242266655, + 0.00937698781490326, + 0.01156091783195734, + 0.007741359528154135, + 0.010631874203681946, + 0.009851336479187012, + 0.010231132619082928, + 0.005781153682619333, + 0.006680578924715519, + 0.00333757558837533, + 0.008430387824773788, + 0.0048175714910030365, + 0.0038719254080206156, + 0.01187372487038374, + 0.006936046294867992, + 0.010715333744883537, + 0.03256118670105934, + 0.009269075468182564, + 0.022061394527554512, + 0.011392684653401375, + 0.01128785964101553, + 0.016842937096953392, + 0.014780852012336254, + 0.006947825197130442, + 0.021758709102869034, + 0.005628929939121008, + 0.013284680433571339, + 0.0071145580150187016, + 0.005829116329550743, + 0.010332939215004444, + 0.006341471802443266, + 0.006293314974755049, + 0.008981875143945217, + 0.005025485996156931, + 0.00861432310193777, + 0.003156013786792755, + 0.005937257315963507, + 0.005499943159520626, + 0.0053738756105303764, + 0.004494717810302973, + 0.0056756106205284595, + 0.003458675928413868, + 0.005258889868855476, + 0.002325913868844509, + 0.004282123874872923, + 0.002851031720638275, + 0.004210026003420353, + 0.002613805001601577, + 0.0025624444242566824, + 0.0021622662898153067, + 0.004041077569127083, + 0.002496646484360099, + 0.002385222353041172, + 0.0017147029284387827, + 0.0031909265089780092, + 0.0040154829621315, + 0.006172850728034973, + 0.002971620298922062, + 0.005004706792533398, + 0.0023161335848271847, + 0.0037744028959423304, + 0.0035886592231690884, + 0.0051024118438363075, + 0.0081639951094985, + 0.011611553840339184, + 0.009268693625926971, + 0.018864845857024193, + 0.006472872104495764, + 0.011534330435097218, + 0.014211868867278099, + 0.028379911556839943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.04305654764175415 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.026344453915953636 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.026344453915953636, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344453915953636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3398197293281555, + 0.12211812287569046, + 0.058845870196819305, + 0.09591792523860931, + 0.04735642299056053, + 0.06381629407405853, + 0.12282135337591171, + 0.0649472028017044, + 0.04900527745485306, + 0.03535185754299164 + ], + "credit_map": [ + 0.03844492882490158, + 0.05155608803033829, + 0.03487967327237129, + 0.050906404852867126, + 0.03440438583493233, + 0.033515579998493195, + 0.04326425492763519, + 0.05284835770726204, + 0.017371397465467453, + 0.0196068175137043, + 0.012513302266597748, + 0.01759023219347, + 0.0116844167932868, + 0.015398737043142319, + 0.013618684373795986, + 0.014334520325064659, + 0.006854623090475798, + 0.008817997761070728, + 0.004790907725691795, + 0.008443095721304417, + 0.004551210906356573, + 0.005364304408431053, + 0.01172312255948782, + 0.008300600573420525, + 0.011128994636237621, + 0.012206929735839367, + 0.01238242443650961, + 0.013426884077489376, + 0.012472049333155155, + 0.01169663667678833, + 0.010106711648404598, + 0.01249728910624981, + 0.006191863212734461, + 0.007630783598870039, + 0.006796434987336397, + 0.0069872005842626095, + 0.0038098415825515985, + 0.0050129215233027935, + 0.004918080288916826, + 0.006009296514093876, + 0.00785643421113491, + 0.005493715405464172, + 0.006980800069868565, + 0.010560239665210247, + 0.005737752187997103, + 0.009485275484621525, + 0.007438572123646736, + 0.010263503529131413, + 0.018437430262565613, + 0.012204959988594055, + 0.014265929348766804, + 0.020960312336683273, + 0.01126612164080143, + 0.016230370849370956, + 0.01098527479916811, + 0.018470941111445427, + 0.00702219782397151, + 0.008658728562295437, + 0.0034422932658344507, + 0.009763228707015514, + 0.00527877127751708, + 0.007760830223560333, + 0.012219744734466076, + 0.010801400989294052, + 0.005992918275296688, + 0.005580536555498838, + 0.004893769975751638, + 0.008345198817551136, + 0.005698696710169315, + 0.006769438274204731, + 0.004110301844775677, + 0.007614410016685724, + 0.0035121599212288857, + 0.0061705042608082294, + 0.002753813285380602, + 0.005129276309162378, + 0.0029594453517347574, + 0.003394122002646327, + 0.0056117624044418335, + 0.005820769350975752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.026344453915953636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from behind and slightly above, captures a man standing on the balcony of a modern high-rise apartment building. He wears a soft, dark gray cotton hoodie and leans slightly forward, his hands resting naturally on the cool black metal railing. His short hair ruffles gently in the evening breeze, and our gaze drifts over his shoulder, focused on his intense gaze. Below, a busy street buzzes with traffic. At dusk, countless headlights and taillights converge into flowing rivers of light, while miniature pedestrians hurried across the sidewalk. The last rays of sunset paint the sky with warm orange-purple hues, contrasting sharply with the gradually brightening cool lights of the city buildings. The complex light highlights the stern profile of his face and the fine textures of his clothing. The entire image has a photographic quality, and the strong depth of field isolates his quiet contemplation from the bustling city below, creating an atmosphere of urban contemplation that is both detached and profound.", + "seed": 23649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_835_seed_23649_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": -0.023699206043966116, + "vlm_logprob": -0.023699206043966116, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.023699206043966116, + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.03305290266871452 + ], + "opposite_logprob_tokens": [ + -3.783052921295166 + ], + "logprob_reward": -0.03305290266871452, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03305290266871452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12088745832443237, + 0.10741869360208511, + 0.09276142716407776, + 0.043750446289777756, + 0.028991304337978363, + 0.04314309358596802, + 0.41844621300697327, + 0.045320481061935425, + 0.05745583400130272, + 0.041824981570243835 + ], + "credit_map": [ + 0.015143334865570068, + 0.016364311799407005, + 0.010086966678500175, + 0.017173592001199722, + 0.01212789211422205, + 0.011347996070981026, + 0.020058466121554375, + 0.01858491078019142, + 0.02844160981476307, + 0.012096868827939034, + 0.009462575428187847, + 0.01181087363511324, + 0.006649510934948921, + 0.007887761108577251, + 0.011682005599141121, + 0.019387507811188698, + 0.00858868844807148, + 0.00616501085460186, + 0.011009931564331055, + 0.02763611264526844, + 0.005283144302666187, + 0.004787789192050695, + 0.008778291754424572, + 0.020512470975518227, + 0.004574475344270468, + 0.005372334271669388, + 0.004575274884700775, + 0.005159009713679552, + 0.0038947921711951494, + 0.005472266580909491, + 0.008142962120473385, + 0.006559336557984352, + 0.005417360924184322, + 0.00358498259447515, + 0.003982685972005129, + 0.0029930153395980597, + 0.0026449246797710657, + 0.0028894308488816023, + 0.0037767263129353523, + 0.0037021790631115437, + 0.005891782231628895, + 0.0038120499812066555, + 0.007232527248561382, + 0.004947264213114977, + 0.006035896018147469, + 0.0027344482950866222, + 0.0047226035967469215, + 0.007766526658087969, + 0.08207092434167862, + 0.022488867864012718, + 0.03723277524113655, + 0.12160447984933853, + 0.01773628406226635, + 0.007627440616488457, + 0.07290130853652954, + 0.05678416043519974, + 0.005412633065134287, + 0.002875580918043852, + 0.011693940497934818, + 0.006658625788986683, + 0.0034480455797165632, + 0.00330531457439065, + 0.0067580086179077625, + 0.005168337840586901, + 0.010493976064026356, + 0.006417868658900261, + 0.0077901738695800304, + 0.009943915531039238, + 0.004742972087115049, + 0.0054799518547952175, + 0.00461658975109458, + 0.007970391772687435, + 0.005463083740323782, + 0.010787351988255978, + 0.0059242043644189835, + 0.006581299006938934, + 0.003326292848214507, + 0.0033939313143491745, + 0.0031195201445370913, + 0.0032293000258505344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.03305290266871452 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.033734291791915894 + ], + "opposite_logprob_tokens": [ + -3.4087343215942383 + ], + "logprob_reward": -0.033734291791915894, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033734291791915894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12815257906913757, + 0.08511954545974731, + 0.16095827519893646, + 0.06743068248033524, + 0.048045575618743896, + 0.04565223306417465, + 0.11265283077955246, + 0.10073506832122803, + 0.19803431630134583, + 0.053218889981508255 + ], + "credit_map": [ + 0.012941407971084118, + 0.01132384967058897, + 0.008894597180187702, + 0.027974290773272514, + 0.012345904484391212, + 0.013622243888676167, + 0.020798970013856888, + 0.020251313224434853, + 0.017594948410987854, + 0.007233296521008015, + 0.007815971039235592, + 0.008470829576253891, + 0.00769426953047514, + 0.006610095500946045, + 0.014954766258597374, + 0.014745371416211128, + 0.015623214654624462, + 0.01262566726654768, + 0.0202980637550354, + 0.0475485660135746, + 0.009332100860774517, + 0.008532225154340267, + 0.012648510746657848, + 0.03434991464018822, + 0.007047481834888458, + 0.011642834171652794, + 0.00930920708924532, + 0.01063455082476139, + 0.006847269367426634, + 0.00478660874068737, + 0.008303346112370491, + 0.008859386667609215, + 0.008012654259800911, + 0.004836053587496281, + 0.006783205550163984, + 0.004846131894737482, + 0.006107652559876442, + 0.0044831992127001286, + 0.006502488628029823, + 0.006474187131971121, + 0.006490768864750862, + 0.005329226143658161, + 0.005739669315516949, + 0.005890933331102133, + 0.005410424899309874, + 0.0029725811909884214, + 0.005042524542659521, + 0.008776105009019375, + 0.021441856399178505, + 0.0063887834548950195, + 0.00795734766870737, + 0.02802177332341671, + 0.00490109808743, + 0.0022030810359865427, + 0.022170957177877426, + 0.019567936658859253, + 0.014618515968322754, + 0.0071944547817111015, + 0.017288506031036377, + 0.012135704979300499, + 0.00786122027784586, + 0.0058361138217151165, + 0.021042292937636375, + 0.014758250676095486, + 0.04923351854085922, + 0.017896270379424095, + 0.02450147457420826, + 0.030168039724230766, + 0.01925050839781761, + 0.01140874158591032, + 0.026130972430109978, + 0.019444776698946953, + 0.0068292212672531605, + 0.0102249039337039, + 0.006189380306750536, + 0.009246603585779667, + 0.005495491437613964, + 0.0037735537625849247, + 0.005648486781865358, + 0.005811244249343872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.033734291791915894 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.0019421778852120042 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019421778852120042, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019421778852120042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22142677009105682, + 0.05374637246131897, + 0.30528146028518677, + 0.0516250915825367, + 0.027214979752898216, + 0.046035606414079666, + 0.1408577263355255, + 0.08126486092805862, + 0.055371060967445374, + 0.0171760693192482 + ], + "credit_map": [ + 0.021219026297330856, + 0.014047535136342049, + 0.0509149469435215, + 0.041990626603364944, + 0.012411803007125854, + 0.016713233664631844, + 0.017305824905633926, + 0.04682377725839615, + 0.006981631275266409, + 0.004448354709893465, + 0.006301491986960173, + 0.010014403611421585, + 0.004053500946611166, + 0.0050797369331121445, + 0.008270224556326866, + 0.008597028441727161, + 0.02860216796398163, + 0.019445031881332397, + 0.03873133659362793, + 0.09511872380971909, + 0.016923390328884125, + 0.015102524310350418, + 0.022396283224225044, + 0.06896200031042099, + 0.007666320074349642, + 0.008294080384075642, + 0.007299651857465506, + 0.006200087256729603, + 0.00565678533166647, + 0.0035552959889173508, + 0.006154099944978952, + 0.006798769813030958, + 0.003637975314632058, + 0.002940213307738304, + 0.0028407343197613955, + 0.0025478992611169815, + 0.0032970160245895386, + 0.003572433954104781, + 0.0040769618935883045, + 0.004301746375858784, + 0.007225779350847006, + 0.004738877061754465, + 0.006325161084532738, + 0.0050031039863824844, + 0.005502132698893547, + 0.0035148835740983486, + 0.004667091649025679, + 0.009058577008545399, + 0.02155693620443344, + 0.007993729785084724, + 0.013352679088711739, + 0.042534343898296356, + 0.006478664930909872, + 0.00262945587746799, + 0.0288263987749815, + 0.017485519871115685, + 0.008755488321185112, + 0.0054510412737727165, + 0.0175878144800663, + 0.011414091102778912, + 0.007081027142703533, + 0.008917621336877346, + 0.013187327422201633, + 0.008870444260537624, + 0.0156712643802166, + 0.004140531178563833, + 0.006530028767883778, + 0.0054090591147542, + 0.0038093295879662037, + 0.003213412594050169, + 0.010235283523797989, + 0.006362149957567453, + 0.0021616278681904078, + 0.0032374172005802393, + 0.002044429536908865, + 0.00356066832318902, + 0.001851914799772203, + 0.0011145664611831307, + 0.001413269666954875, + 0.0017921749968081713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0019421778852120042 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.0336940661072731 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.0336940661072731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336940661072731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11399274319410324, + 0.06538791954517365, + 0.4037967920303345, + 0.019525622949004173, + 0.019443131983280182, + 0.021517161279916763, + 0.1956256777048111, + 0.04348169267177582, + 0.05536125600337982, + 0.06186797469854355 + ], + "credit_map": [ + 0.011095773428678513, + 0.009244619868695736, + 0.02101088874042034, + 0.02063279040157795, + 0.010607581585645676, + 0.01083846390247345, + 0.010411703959107399, + 0.020150916650891304, + 0.011921939440071583, + 0.004409200046211481, + 0.007118527777493, + 0.007763814181089401, + 0.007456941530108452, + 0.00786540936678648, + 0.007962659932672977, + 0.010889428667724133, + 0.03871151804924011, + 0.02527100034058094, + 0.050595201551914215, + 0.12671087682247162, + 0.022062212228775024, + 0.019729910418391228, + 0.02809438481926918, + 0.09262169152498245, + 0.0021529991645365953, + 0.002690089400857687, + 0.002435525180771947, + 0.002044844441115856, + 0.0019784204196184874, + 0.0013771391240879893, + 0.003865511855110526, + 0.0029810939449816942, + 0.002461886266246438, + 0.0032547393348068, + 0.0020424809772521257, + 0.0019811675883829594, + 0.0025254145730286837, + 0.001607480808161199, + 0.002966824686154723, + 0.0026031394954770803, + 0.003980554174631834, + 0.0034448045771569014, + 0.002093345858156681, + 0.00217361724935472, + 0.0030425593722611666, + 0.0014258332084864378, + 0.002119807293638587, + 0.0032366395462304354, + 0.03734787181019783, + 0.0111123351380229, + 0.018954632803797722, + 0.06119200959801674, + 0.008831541985273361, + 0.004280651453882456, + 0.03199516609311104, + 0.021911466494202614, + 0.0044859289191663265, + 0.0028564983513206244, + 0.008745403960347176, + 0.004806913435459137, + 0.004285728558897972, + 0.003534049727022648, + 0.00989732425659895, + 0.004869845230132341, + 0.010590861551463604, + 0.004364070016890764, + 0.006595888175070286, + 0.007047697901725769, + 0.0058581531047821045, + 0.00391981890425086, + 0.008665133267641068, + 0.008319634012877941, + 0.0071455202996730804, + 0.009356527589261532, + 0.005362143274396658, + 0.010427514091134071, + 0.007738989777863026, + 0.004733463749289513, + 0.009273574687540531, + 0.007830244489014149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0336940661072731 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.01607259176671505 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607259176671505, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607259176671505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09205907583236694, + 0.5062888860702515, + 0.17526206374168396, + 0.02427322044968605, + 0.0366019643843174, + 0.019844191148877144, + 0.09919892996549606, + 0.01532035507261753, + 0.01685323193669319, + 0.014298087917268276 + ], + "credit_map": [ + 0.014322252944111824, + 0.010263526812195778, + 0.008930339477956295, + 0.019492704421281815, + 0.00755844172090292, + 0.007112822961062193, + 0.010458440519869328, + 0.013920544646680355, + 0.14341454207897186, + 0.06029711663722992, + 0.04629451781511307, + 0.060968995094299316, + 0.02565201185643673, + 0.038039788603782654, + 0.041977833956480026, + 0.08964407444000244, + 0.016745157539844513, + 0.011493699625134468, + 0.022075679153203964, + 0.05466499552130699, + 0.00979242380708456, + 0.008440864272415638, + 0.012355719693005085, + 0.03969351202249527, + 0.003669047262519598, + 0.004068811424076557, + 0.0037572358269244432, + 0.0025765199679881334, + 0.0019626368302851915, + 0.0022217407822608948, + 0.002899589715525508, + 0.0031176386401057243, + 0.005540826823562384, + 0.005554509349167347, + 0.004842668306082487, + 0.004023060668259859, + 0.003020426258444786, + 0.004326970782130957, + 0.004388302098959684, + 0.004905198700726032, + 0.002595487516373396, + 0.0022013424895703793, + 0.0017014690674841404, + 0.0027312184683978558, + 0.0028208130970597267, + 0.0019926337990909815, + 0.002570691052824259, + 0.0032305351924151182, + 0.013107795268297195, + 0.005894631613045931, + 0.009817620739340782, + 0.03147858753800392, + 0.004969548434019089, + 0.002024916699156165, + 0.02098848484456539, + 0.01091734692454338, + 0.0013535305624827743, + 0.0009769103489816189, + 0.002589813433587551, + 0.0025524827651679516, + 0.0019470821134746075, + 0.0013984886463731527, + 0.0025092698633670807, + 0.001992777455598116, + 0.002362625440582633, + 0.0021594809368252754, + 0.0029584746807813644, + 0.003180010011419654, + 0.0014296346344053745, + 0.0013769568176940084, + 0.0017875067424029112, + 0.0015985430218279362, + 0.0019495963351801038, + 0.00243175495415926, + 0.0013656917726621032, + 0.0014522355049848557, + 0.0017141076968982816, + 0.0013505901442840695, + 0.0019360395381227136, + 0.0020980718545615673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.01607259176671505 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from behind and slightly above, captures a man standing on the balcony of a modern high-rise apartment building. He wears a soft, dark gray cotton hoodie and leans slightly forward, his hands resting naturally on the cool black metal railing. His short hair ruffles gently in the evening breeze, and our gaze drifts over his shoulder, focused on his intense gaze. Below, a busy street buzzes with traffic. At dusk, countless headlights and taillights converge into flowing rivers of light, while miniature pedestrians hurried across the sidewalk. The last rays of sunset paint the sky with warm orange-purple hues, contrasting sharply with the gradually brightening cool lights of the city buildings. The complex light highlights the stern profile of his face and the fine textures of his clothing. The entire image has a photographic quality, and the strong depth of field isolates his quiet contemplation from the bustling city below, creating an atmosphere of urban contemplation that is both detached and profound.", + "seed": 23650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_835_seed_23650_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": -0.06378591633401812, + "vlm_logprob": -0.06378591633401812, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06378591633401812, + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.022807423025369644 + ], + "opposite_logprob_tokens": [ + -4.147807598114014 + ], + "logprob_reward": -0.022807423025369644, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022807423025369644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14166778326034546, + 0.6500635147094727, + 0.03995322436094284, + 0.0928790345788002, + 0.019519347697496414, + 0.012210414744913578, + 0.005584914702922106, + 0.009672718122601509, + 0.017433486878871918, + 0.0110155139118433 + ], + "credit_map": [ + 0.01692359708249569, + 0.010252394713461399, + 0.02021736465394497, + 0.0500703826546669, + 0.005643393844366074, + 0.00644572963938117, + 0.020973049104213715, + 0.011141880415380001, + 0.061611562967300415, + 0.03853736072778702, + 0.07942472398281097, + 0.2186339795589447, + 0.029116474092006683, + 0.022982893511652946, + 0.14750607311725616, + 0.052250511944293976, + 0.004567962605506182, + 0.004579742904752493, + 0.0048365965485572815, + 0.007276179734617472, + 0.0017780369380488992, + 0.006697901524603367, + 0.004495606757700443, + 0.0057212007232010365, + 0.008049950003623962, + 0.005961919203400612, + 0.010904845781624317, + 0.02881808392703533, + 0.005135565530508757, + 0.006870501209050417, + 0.019628582522273064, + 0.007509590592235327, + 0.001759361824952066, + 0.0015764618292450905, + 0.001836521434597671, + 0.003389354096725583, + 0.0011421567760407925, + 0.0028361689765006304, + 0.0033227531239390373, + 0.0036565703339874744, + 0.0009954423876479268, + 0.0013198003871366382, + 0.0014085626462474465, + 0.0026150583289563656, + 0.000816438056062907, + 0.0011535331141203642, + 0.002499878406524658, + 0.001401703106239438, + 0.0006915548583492637, + 0.0006943015032447875, + 0.0007010831031948328, + 0.0009426014148630202, + 0.000393375608837232, + 0.0008527123718522489, + 0.0005353577435016632, + 0.0007739287684671581, + 0.0007262239814735949, + 0.0008364402456209064, + 0.0010202862322330475, + 0.001903028110973537, + 0.0007942821830511093, + 0.0011865957640111446, + 0.0016046250239014626, + 0.0016012374544516206, + 0.003057468682527542, + 0.0012571071274578571, + 0.0024479145649820566, + 0.003988727927207947, + 0.0011987759498879313, + 0.0021707008127123117, + 0.0016453382559120655, + 0.0016674543730914593, + 0.0010484206723049283, + 0.0009136394364759326, + 0.0011762019712477922, + 0.002417580923065543, + 0.000738464412279427, + 0.0004701134457718581, + 0.0031444579362869263, + 0.0011066367151215672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.022807423025369644 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.04308726266026497 + ], + "opposite_logprob_tokens": [ + -3.1680872440338135 + ], + "logprob_reward": -0.04308726266026497, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308726266026497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3147934377193451, + 0.29402828216552734, + 0.12185274064540863, + 0.07967323064804077, + 0.07982658594846725, + 0.020303864032030106, + 0.02510841377079487, + 0.012010646052658558, + 0.028692197054624557, + 0.023710599169135094 + ], + "credit_map": [ + 0.034575205296278, + 0.022179562598466873, + 0.043762363493442535, + 0.09736379981040955, + 0.019071443006396294, + 0.015704328194260597, + 0.05756860971450806, + 0.02456812746822834, + 0.027703791856765747, + 0.0175215695053339, + 0.03470013663172722, + 0.09336390346288681, + 0.01383647508919239, + 0.01167997345328331, + 0.06807180494070053, + 0.027150623500347137, + 0.02461489662528038, + 0.01645163632929325, + 0.01504282746464014, + 0.031016182154417038, + 0.005385525058954954, + 0.006756477057933807, + 0.013066140934824944, + 0.009519063867628574, + 0.012517633847892284, + 0.00635809451341629, + 0.011583469808101654, + 0.02610958181321621, + 0.003654862754046917, + 0.003741850843653083, + 0.010214952751994133, + 0.005492782220244408, + 0.007321901619434357, + 0.010791165754199028, + 0.007155963685363531, + 0.014049289748072624, + 0.005065665114670992, + 0.010736190713942051, + 0.01289780717343092, + 0.011808598414063454, + 0.0021513786632567644, + 0.0023421882651746273, + 0.002339892089366913, + 0.003686791518703103, + 0.0018314222106710076, + 0.0026494739577174187, + 0.0024928029160946608, + 0.0028099133633077145, + 0.002323308028280735, + 0.003265454201027751, + 0.0022150292061269283, + 0.003861907869577408, + 0.0019796600099653006, + 0.0034698094241321087, + 0.004269429948180914, + 0.003723816480487585, + 0.001144849113188684, + 0.0018022237345576286, + 0.0013296183897182345, + 0.0023224474862217903, + 0.0007974794716574252, + 0.0013007165398448706, + 0.0014170617796480656, + 0.0018962500616908073, + 0.005026884842664003, + 0.0027780267409980297, + 0.00472233584150672, + 0.006829126738011837, + 0.0015088585205376148, + 0.0032404609955847263, + 0.0024210878182202578, + 0.002165414160117507, + 0.002312828553840518, + 0.003220020094886422, + 0.003757528029382229, + 0.0050962152890861034, + 0.0013512740842998028, + 0.0013030917616561055, + 0.004185754340142012, + 0.002483886433765292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04308726266026497 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.0028186615090817213 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.0028186615090817213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028186615090817213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2488556206226349, + 0.1680023968219757, + 0.06662113219499588, + 0.2524212598800659, + 0.08956970274448395, + 0.019389908760786057, + 0.02405085600912571, + 0.03291530907154083, + 0.07014284282922745, + 0.02803102321922779 + ], + "credit_map": [ + 0.03123144805431366, + 0.016834856942296028, + 0.029453285038471222, + 0.06767185032367706, + 0.015268598683178425, + 0.020160967484116554, + 0.04538958892226219, + 0.022844994440674782, + 0.01866300031542778, + 0.011318942531943321, + 0.021986406296491623, + 0.05803387984633446, + 0.007295507472008467, + 0.0060536861419677734, + 0.032949164509773254, + 0.011701787821948528, + 0.0066262525506317616, + 0.006499973125755787, + 0.0071138241328299046, + 0.009680403396487236, + 0.004559318069368601, + 0.009903996251523495, + 0.014140761457383633, + 0.008096597157418728, + 0.022439585998654366, + 0.015064810402691364, + 0.029335886240005493, + 0.07714205235242844, + 0.014918060973286629, + 0.010076300241053104, + 0.064417764544487, + 0.01902678608894348, + 0.006864981725811958, + 0.005574348382651806, + 0.008254180662333965, + 0.012697430327534676, + 0.0049425335600972176, + 0.017785202711820602, + 0.012189121916890144, + 0.021261896938085556, + 0.002135177841410041, + 0.001674980390816927, + 0.002323663095012307, + 0.0037327338941395283, + 0.0018659444758668542, + 0.0022455535363405943, + 0.0029047373682260513, + 0.0025071161799132824, + 0.002154500223696232, + 0.0021726060658693314, + 0.0024051114451140165, + 0.003493508091196418, + 0.002122689737007022, + 0.004382495768368244, + 0.0027717126067727804, + 0.004548229742795229, + 0.003331584855914116, + 0.003094050334766507, + 0.0043753269128501415, + 0.00548129016533494, + 0.0031178630888462067, + 0.005059090908616781, + 0.003098354209214449, + 0.005357744637876749, + 0.013227008283138275, + 0.0051612830720841885, + 0.008474402129650116, + 0.015620957128703594, + 0.004699580371379852, + 0.008725260384380817, + 0.006079287733882666, + 0.00815505813807249, + 0.004410331137478352, + 0.002454969333484769, + 0.0027162337210029364, + 0.0037180499639362097, + 0.0017657483695074916, + 0.00313975615426898, + 0.005857471376657486, + 0.003968461882323027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0028186615090817213 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.20155736804008484 + ], + "opposite_logprob_tokens": [ + -1.7015573978424072 + ], + "logprob_reward": -0.20155736804008484, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155736804008484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.247391939163208, + 0.3654628098011017, + 0.08966396003961563, + 0.21002960205078125, + 0.03020932339131832, + 0.01331019401550293, + 0.020717110484838486, + 0.008567324839532375, + 0.00651852460578084, + 0.008129171095788479 + ], + "credit_map": [ + 0.03298351913690567, + 0.01808304525911808, + 0.033967968076467514, + 0.0904601439833641, + 0.00966555718332529, + 0.00804068148136139, + 0.03640412166714668, + 0.017786916345357895, + 0.03378527984023094, + 0.02346659265458584, + 0.04467976838350296, + 0.12235191464424133, + 0.01635853946208954, + 0.014290931634604931, + 0.08212973177433014, + 0.028400039300322533, + 0.015357878059148788, + 0.012484763748943806, + 0.009982551448047161, + 0.013695327565073967, + 0.004733906127512455, + 0.004209085367619991, + 0.021305955946445465, + 0.007894492708146572, + 0.020190859213471413, + 0.013621939346194267, + 0.025559505447745323, + 0.06845963001251221, + 0.010928977280855179, + 0.01331596914678812, + 0.04170050472021103, + 0.016252240166068077, + 0.0022435467690229416, + 0.003350216895341873, + 0.0029326332733035088, + 0.004964078310877085, + 0.0017366345273330808, + 0.005164037924259901, + 0.0039683375507593155, + 0.005849836394190788, + 0.001054689404554665, + 0.001559477299451828, + 0.0011844875989481807, + 0.0027975288685411215, + 0.0010120291262865067, + 0.001847089035436511, + 0.0021919396240264177, + 0.0016629536403343081, + 0.002002882305532694, + 0.0020459310617297888, + 0.0020816745236516, + 0.003328236984089017, + 0.001594238099642098, + 0.0031360387802124023, + 0.0035134670324623585, + 0.0030146425124257803, + 0.0009882643353193998, + 0.0008505662553943694, + 0.0007572566391900182, + 0.0014029251178726554, + 0.0006842143484391272, + 0.0008982257568277419, + 0.0016333971871063113, + 0.001352475257590413, + 0.0008943841676227748, + 0.0008911382756195962, + 0.0008715910953469574, + 0.0011966378660872579, + 0.0004227966710459441, + 0.0006586684612557292, + 0.0007311506778933108, + 0.0008521574200130999, + 0.0006382702849805355, + 0.0006269909790717065, + 0.0007520317449234426, + 0.0016488066175952554, + 0.0005255729192867875, + 0.0008466602303087711, + 0.0017389040440320969, + 0.0013519347412511706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.20155736804008484 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.04865886643528938 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865886643528938, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865886643528938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47921323776245117, + 0.09780057519674301, + 0.20144732296466827, + 0.06860095262527466, + 0.044151671230793, + 0.030241118744015694, + 0.017115941271185875, + 0.024684401229023933, + 0.02736867591738701, + 0.009376058354973793 + ], + "credit_map": [ + 0.05716295540332794, + 0.03580465167760849, + 0.07005754858255386, + 0.17422503232955933, + 0.020156174898147583, + 0.021578166633844376, + 0.0630984976887703, + 0.03713022917509079, + 0.012445710599422455, + 0.007670643739402294, + 0.013362710364162922, + 0.03326663374900818, + 0.0043921275064349174, + 0.00355741661041975, + 0.016568152233958244, + 0.006537177134305239, + 0.016963385045528412, + 0.013073556125164032, + 0.021549314260482788, + 0.06075523793697357, + 0.010707705281674862, + 0.01319101545959711, + 0.04877027869224548, + 0.016436830163002014, + 0.005691094323992729, + 0.004890806972980499, + 0.006954666692763567, + 0.018288232386112213, + 0.005014201160520315, + 0.004767314996570349, + 0.015505744144320488, + 0.007488892879337072, + 0.003531812923029065, + 0.0027089156210422516, + 0.0038827352691441774, + 0.008173182606697083, + 0.003066639881581068, + 0.00641049025580287, + 0.008482987061142921, + 0.007894910871982574, + 0.003125912044197321, + 0.0028167027048766613, + 0.003322531236335635, + 0.004320495296269655, + 0.00355031481012702, + 0.0046483599580824375, + 0.0036101441364735365, + 0.00484665809199214, + 0.0021078954450786114, + 0.0017932261107489467, + 0.001961031462997198, + 0.0031337004620581865, + 0.0011650400701910257, + 0.0018216921016573906, + 0.002745701465755701, + 0.0023876531049609184, + 0.002033211523666978, + 0.0017061954131349921, + 0.0021580944303423166, + 0.004931592848151922, + 0.0019520613132044673, + 0.0028792149387300014, + 0.005455999635159969, + 0.003568031592294574, + 0.002824177499860525, + 0.0021696214098483324, + 0.003401154186576605, + 0.00694166449829936, + 0.0015866558533161879, + 0.0025665152352303267, + 0.005164986476302147, + 0.0027138993609696627, + 0.0009815557859838009, + 0.0013357159914448857, + 0.001793597941286862, + 0.0015267786802724004, + 0.0006876403349451721, + 0.000955368101131171, + 0.0010995552875101566, + 0.0009958455339074135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04865886643528938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from behind and slightly above, captures a man standing on the balcony of a modern high-rise apartment building. He wears a soft, dark gray cotton hoodie and leans slightly forward, his hands resting naturally on the cool black metal railing. His short hair ruffles gently in the evening breeze, and our gaze drifts over his shoulder, focused on his intense gaze. Below, a busy street buzzes with traffic. At dusk, countless headlights and taillights converge into flowing rivers of light, while miniature pedestrians hurried across the sidewalk. The last rays of sunset paint the sky with warm orange-purple hues, contrasting sharply with the gradually brightening cool lights of the city buildings. The complex light highlights the stern profile of his face and the fine textures of his clothing. The entire image has a photographic quality, and the strong depth of field isolates his quiet contemplation from the bustling city below, creating an atmosphere of urban contemplation that is both detached and profound.", + "seed": 23651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_4/index_835_seed_23651_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": -0.030390972178429364, + "vlm_logprob": -0.030390972178429364, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.030390972178429364, + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.014436440542340279 + ], + "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": [ + 0.3012545108795166, + 0.13904602825641632, + 0.04763953760266304, + 0.07145244628190994, + 0.03118758276104927, + 0.03843658044934273, + 0.034343454986810684, + 0.05256113037467003, + 0.022789068520069122, + 0.26128965616226196 + ], + "credit_map": [ + 0.029158659279346466, + 0.06006544828414917, + 0.040025483816862106, + 0.05837705358862877, + 0.025412051007151604, + 0.02093118615448475, + 0.0367538258433342, + 0.030530808493494987, + 0.011746397241950035, + 0.03429311141371727, + 0.016703814268112183, + 0.02560446411371231, + 0.01005316898226738, + 0.010244210250675678, + 0.014477284625172615, + 0.015923570841550827, + 0.0049025267362594604, + 0.006211338099092245, + 0.006248605437576771, + 0.005618283059448004, + 0.005509744398295879, + 0.006219093222171068, + 0.0065056756138801575, + 0.006424269638955593, + 0.006835701875388622, + 0.011973370797932148, + 0.010617272928357124, + 0.012755378149449825, + 0.006086369045078754, + 0.00683616241440177, + 0.008545402437448502, + 0.007802794221788645, + 0.002692049602046609, + 0.004323124885559082, + 0.0039602420292794704, + 0.004866330418735743, + 0.0025502692442387342, + 0.0035942085087299347, + 0.00506175123155117, + 0.004139607306569815, + 0.004665536805987358, + 0.006673709489405155, + 0.005929705686867237, + 0.0046468074433505535, + 0.004113506060093641, + 0.004613041877746582, + 0.003459896659478545, + 0.004334377124905586, + 0.003323834855109453, + 0.007304779253900051, + 0.004520568996667862, + 0.005647232290357351, + 0.002957460004836321, + 0.002879309467971325, + 0.004625327419489622, + 0.0030849429313093424, + 0.004971208982169628, + 0.011472723446786404, + 0.005306238308548927, + 0.0066531747579574585, + 0.005192908458411694, + 0.005928257945924997, + 0.007915188558399677, + 0.0051214322447776794, + 0.0022601557429879904, + 0.00404583103954792, + 0.002753198379650712, + 0.00365777057595551, + 0.0018256092444062233, + 0.0016515254974365234, + 0.002911861054599285, + 0.003683118848130107, + 0.03074774704873562, + 0.03700820356607437, + 0.048621442168951035, + 0.036962635815143585, + 0.034322306513786316, + 0.0273837111890316, + 0.025957096368074417, + 0.02028653211891651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "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 shot a medium shot of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.0549742691218853 + ], + "opposite_logprob_tokens": [ + -2.929974317550659 + ], + "logprob_reward": -0.0549742691218853, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549742691218853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42307814955711365, + 0.16834017634391785, + 0.06166011095046997, + 0.0458679124712944, + 0.07323431223630905, + 0.0337417870759964, + 0.025144463405013084, + 0.028352227061986923, + 0.08943387866020203, + 0.05114693567156792 + ], + "credit_map": [ + 0.059923600405454636, + 0.09496515989303589, + 0.047768041491508484, + 0.053273700177669525, + 0.033022768795490265, + 0.049492064863443375, + 0.04073217883706093, + 0.04390065371990204, + 0.013492091558873653, + 0.033485982567071915, + 0.018779903650283813, + 0.03432374820113182, + 0.012383733876049519, + 0.011234128847718239, + 0.023849988356232643, + 0.02079060487449169, + 0.0053564137779176235, + 0.010552593506872654, + 0.008104493841528893, + 0.01230173371732235, + 0.0058488668873906136, + 0.005605220794677734, + 0.007339298725128174, + 0.006551491562277079, + 0.004812780302017927, + 0.008934536948800087, + 0.005395412445068359, + 0.006938600912690163, + 0.004175055306404829, + 0.00356453750282526, + 0.005629958119243383, + 0.0064170327968895435, + 0.007788670249283314, + 0.017687629908323288, + 0.009406369179487228, + 0.010339702479541302, + 0.005068019963800907, + 0.006137356162071228, + 0.0079940902069211, + 0.008812476880848408, + 0.003098130226135254, + 0.007358144968748093, + 0.004409175831824541, + 0.006355382967740297, + 0.0023416411131620407, + 0.0025277091190218925, + 0.004502793308347464, + 0.003148807445541024, + 0.0020906974095851183, + 0.005640642251819372, + 0.003096128348261118, + 0.0037075099535286427, + 0.0024687827099114656, + 0.002072560368105769, + 0.0032133255153894424, + 0.0028548170812427998, + 0.002574252663180232, + 0.008359664119780064, + 0.0031715494114905596, + 0.003987071104347706, + 0.0022839379962533712, + 0.0017320895567536354, + 0.0034006889909505844, + 0.0028429727535694838, + 0.012709269300103188, + 0.016499243676662445, + 0.013990933075547218, + 0.013883354142308235, + 0.0066511076875030994, + 0.009918296709656715, + 0.008309758268296719, + 0.007471921853721142, + 0.006135294213891029, + 0.008689574897289276, + 0.008675703778862953, + 0.008678292855620384, + 0.003810213878750801, + 0.005788201466202736, + 0.004897897131741047, + 0.004471753723919392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0549742691218853 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5549870729446411, + 0.07698975503444672, + 0.039645176380872726, + 0.04442523047327995, + 0.013035094365477562, + 0.011885513551533222, + 0.040405094623565674, + 0.044787224382162094, + 0.10170554369688034, + 0.07213430851697922 + ], + "credit_map": [ + 0.07025507092475891, + 0.1056433692574501, + 0.07108806818723679, + 0.07549416273832321, + 0.050333984196186066, + 0.07332456856966019, + 0.05191001296043396, + 0.05693778395652771, + 0.008051494136452675, + 0.015243693254888058, + 0.00975733995437622, + 0.008749349042773247, + 0.008390957489609718, + 0.009552129544317722, + 0.009753509424626827, + 0.0074912807904183865, + 0.004236677195876837, + 0.007543525192886591, + 0.004506742116063833, + 0.005316257011145353, + 0.003904589917510748, + 0.005601157899945974, + 0.003329524537548423, + 0.005206703208386898, + 0.003918794449418783, + 0.009073392488062382, + 0.005545558873564005, + 0.009907357394695282, + 0.003037927206605673, + 0.003450946416705847, + 0.00500252190977335, + 0.004488731734454632, + 0.0015173695283010602, + 0.0025867342483252287, + 0.0017043035477399826, + 0.0015367475571110845, + 0.0011580019490793347, + 0.0011505180737003684, + 0.0017281522741541266, + 0.001653266604989767, + 0.0011569203343242407, + 0.0017655169358476996, + 0.0016049158293753862, + 0.001615083427168429, + 0.0013943351805210114, + 0.001333490596152842, + 0.0016280212439596653, + 0.001387229422107339, + 0.005019074305891991, + 0.007139483466744423, + 0.005770049523562193, + 0.005005124490708113, + 0.003309311345219612, + 0.0053665670566260815, + 0.004242830444127321, + 0.0045526535250246525, + 0.006060014944523573, + 0.005568370223045349, + 0.006875636056065559, + 0.00524985371157527, + 0.005765929818153381, + 0.005204157903790474, + 0.005544827785342932, + 0.004518431611359119, + 0.012212632223963737, + 0.01602020300924778, + 0.014813685789704323, + 0.015378478914499283, + 0.009260043501853943, + 0.013100224547088146, + 0.011226312257349491, + 0.009693969041109085, + 0.006287673022598028, + 0.009672370739281178, + 0.011656842194497585, + 0.007890193723142147, + 0.014081330969929695, + 0.00769153842702508, + 0.008825795724987984, + 0.006028561387211084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004093957599252462 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.04864955693483353 + ], + "opposite_logprob_tokens": [ + -3.048649549484253 + ], + "logprob_reward": -0.04864955693483353, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864955693483353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3902694880962372, + 0.10928665846586227, + 0.0631072074174881, + 0.07621869444847107, + 0.03943056985735893, + 0.04000724107027054, + 0.06135958060622215, + 0.06977560371160507, + 0.028753986582159996, + 0.12179096788167953 + ], + "credit_map": [ + 0.040014177560806274, + 0.07316712290048599, + 0.037281621247529984, + 0.06933001428842545, + 0.03626612573862076, + 0.0528733991086483, + 0.040113046765327454, + 0.04122402146458626, + 0.007006193045526743, + 0.020874202251434326, + 0.013960219919681549, + 0.01796106994152069, + 0.008042626082897186, + 0.011277645826339722, + 0.018053067848086357, + 0.012111635878682137, + 0.006306702271103859, + 0.011102395132184029, + 0.006717135198414326, + 0.010735442861914635, + 0.00565654831007123, + 0.008338283747434616, + 0.007822957821190357, + 0.006427742075175047, + 0.0081842802464962, + 0.011982222087681293, + 0.014389892108738422, + 0.012836663983762264, + 0.006130438297986984, + 0.005678535904735327, + 0.007737003266811371, + 0.009279660880565643, + 0.004898749757558107, + 0.00705651706084609, + 0.004672889597713947, + 0.006584842689335346, + 0.003985982853919268, + 0.0030257299076765776, + 0.00481831980869174, + 0.00438754353672266, + 0.005040345247834921, + 0.009997030720114708, + 0.005016972776502371, + 0.005137930158525705, + 0.003915443550795317, + 0.0032736919820308685, + 0.003214593045413494, + 0.004411236383020878, + 0.006213175132870674, + 0.016553644090890884, + 0.007302162703126669, + 0.009916127659380436, + 0.004359448794275522, + 0.004764494951814413, + 0.006562604568898678, + 0.005687925964593887, + 0.007033522240817547, + 0.022807173430919647, + 0.0074350256472826, + 0.01029169000685215, + 0.005225365050137043, + 0.004761758726090193, + 0.005230313166975975, + 0.006990752648562193, + 0.004050758667290211, + 0.00370377441868186, + 0.006367654539644718, + 0.00435736496001482, + 0.002123352373018861, + 0.003179528983309865, + 0.0022340097930282354, + 0.002737544709816575, + 0.015389570966362953, + 0.017675606533885002, + 0.024071093648672104, + 0.019059518352150917, + 0.010419828817248344, + 0.013705309480428696, + 0.011829940602183342, + 0.009640107862651348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.04864955693483353 + }, + { + "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 of a man standing on a balcony of a modern high-rise apartment building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, dark gray cotton hoodie and leaning slightly forward with hands on a black metal railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dusk with a sky painted in warm orange-purple hues and flowing rivers of city headlights and taillights below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned behind and slightly above the man, looking over his shoulder at his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a strong depth of field that isolates the man from the bustling city 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": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2588522136211395, + 0.23776668310165405, + 0.06124481186270714, + 0.11646227538585663, + 0.02972262166440487, + 0.023068081587553024, + 0.03850087895989418, + 0.026847464963793755, + 0.13488593697547913, + 0.07264908403158188 + ], + "credit_map": [ + 0.03030095435678959, + 0.035631947219371796, + 0.03519832342863083, + 0.029941795393824577, + 0.03287525847554207, + 0.032813187688589096, + 0.034635990858078, + 0.027454718947410583, + 0.023339632898569107, + 0.024242617189884186, + 0.04533066600561142, + 0.03411059081554413, + 0.021262772381305695, + 0.016228465363383293, + 0.046745043247938156, + 0.026506880298256874, + 0.007594682741910219, + 0.012189404107630253, + 0.012353549711406231, + 0.008456035517156124, + 0.004466423764824867, + 0.004761810414493084, + 0.005097812507301569, + 0.006325085647404194, + 0.012025784701108932, + 0.023104527965188026, + 0.01362734567373991, + 0.025952354073524475, + 0.007434180937707424, + 0.009054508991539478, + 0.014390731230378151, + 0.01087283156812191, + 0.003731116419658065, + 0.00456377025693655, + 0.004351045936346054, + 0.00510743074119091, + 0.002895135199651122, + 0.0028153194580227137, + 0.002888741437345743, + 0.0033700610511004925, + 0.0021030388306826353, + 0.005189412739127874, + 0.0025773814413696527, + 0.0037540700286626816, + 0.0016863411292433739, + 0.0019631132017821074, + 0.0031891311518847942, + 0.0026055900380015373, + 0.0034843571484088898, + 0.0049880617298185825, + 0.0071673085913062096, + 0.005401690490543842, + 0.003981331828981638, + 0.0032736274879425764, + 0.006364472676068544, + 0.0038400294724851847, + 0.002265088725835085, + 0.0036723727826029062, + 0.004276398103684187, + 0.00381841161288321, + 0.0024705466348677874, + 0.002211896236985922, + 0.004970217123627663, + 0.0031625311821699142, + 0.018715202808380127, + 0.02141173928976059, + 0.020809805020689964, + 0.022193046286702156, + 0.010589717887341976, + 0.016017882153391838, + 0.013508321717381477, + 0.011640207841992378, + 0.007537419442087412, + 0.008954695425927639, + 0.013835396617650986, + 0.008483901619911194, + 0.011552555486559868, + 0.007986530661582947, + 0.007509726099669933, + 0.006788851227611303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02980063669383526 + } + ], + "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/0000060/group_5/index_tensor([3553, 3553])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_tensor([3553, 3553])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e3b3d67a680f5bb102fe0940f1d3f557683517e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_tensor([3553, 3553])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, shot from a level angle, captured the historic moment in extreme slow motion. In the frame, a complex lunar lander, clad in gold and silver reflective foil, descended silently and majestically toward the camera. In the airless vacuum of space, the sole source of sunlight cast a sharp light, blinding one side of the lander's metal frame while plunging the other into an abyss of pure darkness. As the lander's circular footpads neared touchdown, its invisible thrusters ejected the fine gray lunar dust from beneath in graceful radial arcs, each particle's trajectory clearly visible as it floated briefly in the weightless environment before cascading back down. Finally, the lander gently and steadily touched the lunar surface, stirring up the final puff of dust that slowly dissipated, and the world fell into absolute silence. Then, an astronaut, clad in a bulky but functional white spacesuit, his helmet covered with gold visors, carefully descended from the lander's ramp. Every movement is rendered with an air of exceptional lightness and poetry by the low gravity and slow motion. When his massive, deeply grooved boot finally touches the soft lunar surface, its sole sinks gently into the powdery lunar soil, leaving a clear, perfect, and eternal footprint. The camera slowly pulls out, framing the astronaut's tiny figure and the lander behind him, revealing the vast, desolate lunar landscape behind them, studded with craters and rocks, stretching to a curved horizon where it meets the inky blackness of the universe. This creates an epic atmosphere of grandeur, loneliness, and human courage.", + "seed": 24616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_3553_seed_24616_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": -0.1745202607475221, + "vlm_logprob": -0.1745202607475221, + "vlm_yes_no_margin": 2.649999964237213, + "vlm_token_credit": -0.1745202607475221, + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5164901614189148 + ], + "opposite_logprob_tokens": [ + -1.0164902210235596 + ], + "logprob_reward": -0.5164901614189148, + "yes_no_margin_reward": 0.5000000596046448, + "token_credit_reward": -0.5164901614189148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06899980455636978, + 0.045254044234752655, + 0.02467503398656845, + 0.03541019558906555, + 0.0600699819624424, + 0.054447147995233536, + 0.10806826502084732, + 0.196306973695755, + 0.22006943821907043, + 0.18669910728931427 + ], + "credit_map": [ + 0.005711252801120281, + 0.005950717255473137, + 0.012265867553651333, + 0.01013246737420559, + 0.007802939508110285, + 0.009050595574080944, + 0.007938600145280361, + 0.010147365741431713, + 0.006184972357004881, + 0.006535045336931944, + 0.0061256904155015945, + 0.003932124935090542, + 0.006755108945071697, + 0.006759341340512037, + 0.004700326826423407, + 0.004261433146893978, + 0.0023362869396805763, + 0.0022823249455541372, + 0.0036956947296857834, + 0.0034246877767145634, + 0.0036279167979955673, + 0.0035925949923694134, + 0.0029855770990252495, + 0.002729950472712517, + 0.00510411337018013, + 0.004162967670708895, + 0.0038169759791344404, + 0.00507043581455946, + 0.00455661304295063, + 0.0047086626291275024, + 0.0035553844645619392, + 0.004435042850673199, + 0.008966471068561077, + 0.006945531349629164, + 0.008113149553537369, + 0.0077195921912789345, + 0.008272076025605202, + 0.007883435115218163, + 0.00801180675625801, + 0.004157923627644777, + 0.005350724793970585, + 0.005053824745118618, + 0.010257040150463581, + 0.00956560019403696, + 0.004229391925036907, + 0.007453774102032185, + 0.008516756817698479, + 0.004020035732537508, + 0.011310815811157227, + 0.011948704719543457, + 0.009868456982076168, + 0.02903895452618599, + 0.01121058315038681, + 0.017673958092927933, + 0.00974614080041647, + 0.007270650006830692, + 0.021308477967977524, + 0.01074343454092741, + 0.02539854496717453, + 0.059320252388715744, + 0.012536934576928616, + 0.03875221684575081, + 0.017071394249796867, + 0.01117571722716093, + 0.016577115282416344, + 0.03576914593577385, + 0.021641334518790245, + 0.03569049388170242, + 0.03281246870756149, + 0.020050382241606712, + 0.03827289864420891, + 0.01925559900701046, + 0.014642994850873947, + 0.022976689040660858, + 0.01784317009150982, + 0.02888304740190506, + 0.01923503540456295, + 0.017569703981280327, + 0.0454382598400116, + 0.020110201090574265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.5164901614189148 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067971679382026196 + ], + "opposite_logprob_tokens": [ + -5.0067973136901855 + ], + "logprob_reward": -0.0067971679382026196, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067971679382026196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11731123924255371, + 0.07510902732610703, + 0.05341588705778122, + 0.056180331856012344, + 0.06301724910736084, + 0.0623118057847023, + 0.09160927683115005, + 0.16324688494205475, + 0.19479051232337952, + 0.12300775200128555 + ], + "credit_map": [ + 0.011194219812750816, + 0.012721569277346134, + 0.024661820381879807, + 0.015243452973663807, + 0.013776060193777084, + 0.015888888388872147, + 0.011646619066596031, + 0.012178605422377586, + 0.012486262246966362, + 0.008312376216053963, + 0.01043422520160675, + 0.009754710830748081, + 0.008622643537819386, + 0.009971776977181435, + 0.009150022640824318, + 0.006377010606229305, + 0.008253723382949829, + 0.0056287203915417194, + 0.006096496246755123, + 0.005746040027588606, + 0.0093165198341012, + 0.008434630930423737, + 0.0056770662777125835, + 0.004262690432369709, + 0.007183903828263283, + 0.008273408748209476, + 0.006758824456483126, + 0.005864844657480717, + 0.008587196469306946, + 0.008036023937165737, + 0.005905614234507084, + 0.005570518784224987, + 0.0056166620925068855, + 0.0063847326673567295, + 0.00579060148447752, + 0.009259097278118134, + 0.00597864156588912, + 0.006626147776842117, + 0.01641734130680561, + 0.006944024004042149, + 0.006362718064337969, + 0.005921604111790657, + 0.012476874515414238, + 0.011371753178536892, + 0.005018320865929127, + 0.010488634929060936, + 0.006647008005529642, + 0.00402489211410284, + 0.009462099522352219, + 0.012428492307662964, + 0.009151067584753036, + 0.023389941081404686, + 0.009588832966983318, + 0.014604907482862473, + 0.006535865366458893, + 0.006448069121688604, + 0.018687797710299492, + 0.009252714924514294, + 0.024102194234728813, + 0.0418088398873806, + 0.013971107080578804, + 0.029816802591085434, + 0.01364164985716343, + 0.011965788900852203, + 0.014923040755093098, + 0.04086440056562424, + 0.022538859397172928, + 0.024070389568805695, + 0.03695646673440933, + 0.023241057991981506, + 0.022008415311574936, + 0.010187886655330658, + 0.010129696689546108, + 0.013209442608058453, + 0.01670853979885578, + 0.019786207005381584, + 0.011686712503433228, + 0.013057575561106205, + 0.026671839877963066, + 0.011757725849747658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0067971679382026196 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018205799162387848 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205799162387848, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018205799162387848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04777713119983673, + 0.03226614370942116, + 0.027173331007361412, + 0.05991293489933014, + 0.04184430465102196, + 0.06008019670844078, + 0.07517334818840027, + 0.3329349458217621, + 0.20428676903247833, + 0.1185508742928505 + ], + "credit_map": [ + 0.005697832442820072, + 0.004161555785685778, + 0.00801075715571642, + 0.008288227953016758, + 0.00615109596401453, + 0.006640894804149866, + 0.0034856130369007587, + 0.005341152660548687, + 0.0037046982906758785, + 0.0031003556214272976, + 0.0065047601237893105, + 0.0033951604273170233, + 0.003475793171674013, + 0.004731852561235428, + 0.003734271740540862, + 0.003619253635406494, + 0.0032365634106099606, + 0.0030810220632702112, + 0.0035270308144390583, + 0.0032518126536160707, + 0.004356901161372662, + 0.003924831282347441, + 0.003099522553384304, + 0.0026956479996442795, + 0.007786155678331852, + 0.006266176234930754, + 0.006347930058836937, + 0.013104049488902092, + 0.0058003501035273075, + 0.007491188123822212, + 0.006165123078972101, + 0.006951960735023022, + 0.005627132020890713, + 0.004230719991028309, + 0.006480686832219362, + 0.004597622435539961, + 0.005310624837875366, + 0.0059834313578903675, + 0.005573770962655544, + 0.004040314815938473, + 0.005303916987031698, + 0.007049233186990023, + 0.011526443995535374, + 0.011751377955079079, + 0.005527188070118427, + 0.007168628741055727, + 0.007889434695243835, + 0.003863973543047905, + 0.006932443007826805, + 0.008552996441721916, + 0.007439234759658575, + 0.019953003153204918, + 0.0075515867210924625, + 0.01302595715969801, + 0.007010210305452347, + 0.0047079212963581085, + 0.038122620433568954, + 0.024608664214611053, + 0.05700043588876724, + 0.0713701993227005, + 0.02892259508371353, + 0.05760475993156433, + 0.02791772596538067, + 0.02738795056939125, + 0.015305927023291588, + 0.03726629167795181, + 0.023071566596627235, + 0.022869247943162918, + 0.03282325714826584, + 0.020957101136446, + 0.035745732486248016, + 0.01624763198196888, + 0.008700278587639332, + 0.014135604724287987, + 0.02090221643447876, + 0.02322199009358883, + 0.010123778134584427, + 0.010173104703426361, + 0.02148745022714138, + 0.00980644952505827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.018205799162387848 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789884552359581 + ], + "opposite_logprob_tokens": [ + -2.578988552093506 + ], + "logprob_reward": -0.0789884552359581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789884552359581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03771243616938591, + 0.02933586947619915, + 0.025186119601130486, + 0.012753334827721119, + 0.022688735276460648, + 0.01824258826673031, + 0.08602656424045563, + 0.1802329570055008, + 0.5382372140884399, + 0.04958416521549225 + ], + "credit_map": [ + 0.004088647663593292, + 0.003987932577729225, + 0.005991154350340366, + 0.005196808837354183, + 0.005351292435079813, + 0.005192109849303961, + 0.003369263606145978, + 0.004535227548331022, + 0.003050463041290641, + 0.0035236766561865807, + 0.006392740644514561, + 0.0034834572579711676, + 0.0029649657662957907, + 0.004352201707661152, + 0.0028481404297053814, + 0.002720222808420658, + 0.00330482330173254, + 0.002712456276640296, + 0.0030304163228720427, + 0.002809205325320363, + 0.004086285829544067, + 0.003806908382102847, + 0.0026435486506670713, + 0.002792477374896407, + 0.0008914289064705372, + 0.001011845888569951, + 0.0017452783649787307, + 0.0030683090444654226, + 0.0008862060494720936, + 0.001700715278275311, + 0.0018319676164537668, + 0.0016175833297893405, + 0.0021130964159965515, + 0.0019371287198737264, + 0.004627601243555546, + 0.0028619742952287197, + 0.0017258304869756103, + 0.0026510905008763075, + 0.00357532175257802, + 0.003196691395714879, + 0.00166863645426929, + 0.0020783189684152603, + 0.004598021507263184, + 0.0020330888219177723, + 0.0013833922566846013, + 0.003531903959810734, + 0.001861433032900095, + 0.0010877957101911306, + 0.008880416862666607, + 0.009741843678057194, + 0.008448970504105091, + 0.023352140560746193, + 0.007717807311564684, + 0.014727458357810974, + 0.007098689209669828, + 0.0060592410154640675, + 0.019718561321496964, + 0.018461570143699646, + 0.0277416929602623, + 0.037095144391059875, + 0.014748378656804562, + 0.028753971680998802, + 0.016629159450531006, + 0.017084475606679916, + 0.03888443484902382, + 0.09869279712438583, + 0.052387017756700516, + 0.07838576287031174, + 0.09304873645305634, + 0.057003702968358994, + 0.08029100298881531, + 0.039543747901916504, + 0.004679417237639427, + 0.006242458708584309, + 0.00643967418000102, + 0.008418367244303226, + 0.006308204960078001, + 0.006019430700689554, + 0.007390920072793961, + 0.004085692577064037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0789884552359581 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521197199821472 + ], + "opposite_logprob_tokens": [ + -1.502119779586792 + ], + "logprob_reward": -0.2521197199821472, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521197199821472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0324154831469059, + 0.02494114264845848, + 0.023922691121697426, + 0.05951271951198578, + 0.02841249108314514, + 0.05158006027340889, + 0.16014373302459717, + 0.4714665412902832, + 0.09184500575065613, + 0.05576009303331375 + ], + "credit_map": [ + 0.0030237280298024416, + 0.0031248319428414106, + 0.006670116446912289, + 0.004736812319606543, + 0.0033857726957648993, + 0.00401316536590457, + 0.0027399680111557245, + 0.0047210874035954475, + 0.0025080135092139244, + 0.0023092678748071194, + 0.004422564525157213, + 0.003999179694801569, + 0.002643160754814744, + 0.003652773331850767, + 0.0026879708748310804, + 0.0027182113844901323, + 0.002755735069513321, + 0.0018579458119347692, + 0.003970813471823931, + 0.0036220441106706858, + 0.0032838264014571905, + 0.003304604673758149, + 0.00279684248380363, + 0.002330879447981715, + 0.007631555665284395, + 0.0069563123397529125, + 0.006723618600517511, + 0.012724255211651325, + 0.005942800547927618, + 0.007401002570986748, + 0.005846273619681597, + 0.006286901421844959, + 0.004809532780200243, + 0.0038688252680003643, + 0.004971472080796957, + 0.002911744173616171, + 0.00275254575535655, + 0.003732881508767605, + 0.00278475321829319, + 0.0025807355996221304, + 0.006085999775677919, + 0.004359815735369921, + 0.011593488976359367, + 0.005081431940197945, + 0.005263939965516329, + 0.009567809291183949, + 0.005397763568907976, + 0.004229810554534197, + 0.01405993290245533, + 0.015083333477377892, + 0.016161751002073288, + 0.04749974235892296, + 0.013280595652759075, + 0.029391618445515633, + 0.012394356541335583, + 0.012272403575479984, + 0.05044908449053764, + 0.03270365670323372, + 0.07324047386646271, + 0.1267782300710678, + 0.03485821187496185, + 0.08405666053295135, + 0.037664104253053665, + 0.031716130673885345, + 0.006686340551823378, + 0.017586108297109604, + 0.009133349172770977, + 0.012017196044325829, + 0.019525663927197456, + 0.010963423177599907, + 0.010369406081736088, + 0.005563512444496155, + 0.005843218415975571, + 0.006147194653749466, + 0.007356192450970411, + 0.008971530012786388, + 0.006726446561515331, + 0.009822250343859196, + 0.007139163091778755, + 0.0037540984340012074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.2521197199821472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a level angle, captured the historic moment in extreme slow motion. In the frame, a complex lunar lander, clad in gold and silver reflective foil, descended silently and majestically toward the camera. In the airless vacuum of space, the sole source of sunlight cast a sharp light, blinding one side of the lander's metal frame while plunging the other into an abyss of pure darkness. As the lander's circular footpads neared touchdown, its invisible thrusters ejected the fine gray lunar dust from beneath in graceful radial arcs, each particle's trajectory clearly visible as it floated briefly in the weightless environment before cascading back down. Finally, the lander gently and steadily touched the lunar surface, stirring up the final puff of dust that slowly dissipated, and the world fell into absolute silence. Then, an astronaut, clad in a bulky but functional white spacesuit, his helmet covered with gold visors, carefully descended from the lander's ramp. Every movement is rendered with an air of exceptional lightness and poetry by the low gravity and slow motion. When his massive, deeply grooved boot finally touches the soft lunar surface, its sole sinks gently into the powdery lunar soil, leaving a clear, perfect, and eternal footprint. The camera slowly pulls out, framing the astronaut's tiny figure and the lander behind him, revealing the vast, desolate lunar landscape behind them, studded with craters and rocks, stretching to a curved horizon where it meets the inky blackness of the universe. This creates an epic atmosphere of grandeur, loneliness, and human courage.", + "seed": 24617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_3553_seed_24617_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": -0.10623885933309793, + "vlm_logprob": -0.10623885933309793, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.10623885933309793, + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09454067796468735 + ], + "opposite_logprob_tokens": [ + -4.219540596008301 + ], + "logprob_reward": -0.09454067796468735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.09454067796468735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15509866178035736, + 0.05429461598396301, + 0.058644820004701614, + 0.06655582785606384, + 0.1586800515651703, + 0.13928230106830597, + 0.028554003685712814, + 0.10827483981847763, + 0.07462822645902634, + 0.15598666667938232 + ], + "credit_map": [ + 0.020548982545733452, + 0.01878349669277668, + 0.01611229218542576, + 0.015724841505289078, + 0.023649288341403008, + 0.02405364252626896, + 0.0211642999202013, + 0.015061821788549423, + 0.00657310988754034, + 0.008107739500701427, + 0.006271726917475462, + 0.005899870302528143, + 0.010106556117534637, + 0.009362703189253807, + 0.0037800411228090525, + 0.00419286685064435, + 0.007980682887136936, + 0.007417020387947559, + 0.009038794785737991, + 0.004219248425215483, + 0.005525646731257439, + 0.008788533508777618, + 0.00939518678933382, + 0.006279706954956055, + 0.006336007732897997, + 0.010474209673702717, + 0.008137251250445843, + 0.005961824208498001, + 0.007937747053802013, + 0.008064355701208115, + 0.013349813409149647, + 0.006294624879956245, + 0.018668677657842636, + 0.02618670091032982, + 0.02164958417415619, + 0.015515838749706745, + 0.024427542462944984, + 0.020352773368358612, + 0.0182968620210886, + 0.013582063838839531, + 0.020167334005236626, + 0.014766378328204155, + 0.013296212069690228, + 0.012819020077586174, + 0.034652844071388245, + 0.02257554791867733, + 0.013993369415402412, + 0.007011591922491789, + 0.00379729806445539, + 0.0035196675453335047, + 0.0028490088880062103, + 0.002355974167585373, + 0.004023998975753784, + 0.004385023843497038, + 0.004980427213013172, + 0.0026426054537296295, + 0.013689259067177773, + 0.019536130130290985, + 0.0122667346149683, + 0.009078658185899258, + 0.016114823520183563, + 0.01980244368314743, + 0.010422413237392902, + 0.007364374585449696, + 0.007320598233491182, + 0.011854945681989193, + 0.005439505446702242, + 0.005718652158975601, + 0.013084180653095245, + 0.01516530942171812, + 0.008777569979429245, + 0.007267465349286795, + 0.013025954365730286, + 0.02656010538339615, + 0.02785450406372547, + 0.01629221998155117, + 0.02530515380203724, + 0.011523829773068428, + 0.01801617629826069, + 0.01740872673690319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.09454067796468735 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983754687011242 + ], + "opposite_logprob_tokens": [ + -3.5298376083374023 + ], + "logprob_reward": -0.02983754687011242, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983754687011242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16767774522304535, + 0.09684152901172638, + 0.051399149000644684, + 0.05737663432955742, + 0.2841680943965912, + 0.04996563866734505, + 0.031002873554825783, + 0.09012673795223236, + 0.0876091793179512, + 0.08383242040872574 + ], + "credit_map": [ + 0.017431648448109627, + 0.019816137850284576, + 0.016836419701576233, + 0.012218168936669827, + 0.036126889288425446, + 0.026781652122735977, + 0.020474309101700783, + 0.01799250766634941, + 0.01543118804693222, + 0.013232233002781868, + 0.011647244915366173, + 0.00788225419819355, + 0.019888613373041153, + 0.015798993408679962, + 0.006782643496990204, + 0.006178362760692835, + 0.007338013965636492, + 0.005347525700926781, + 0.0079099852591753, + 0.003858650103211403, + 0.005102874711155891, + 0.00747399078682065, + 0.008335314691066742, + 0.006032791920006275, + 0.005841445177793503, + 0.0072557865642011166, + 0.0055239531211555, + 0.005529256071895361, + 0.006158663425594568, + 0.006552238017320633, + 0.014239054173231125, + 0.006276239175349474, + 0.03465600684285164, + 0.07443682104349136, + 0.019533827900886536, + 0.02608213946223259, + 0.043944310396909714, + 0.04462958872318268, + 0.02446633391082287, + 0.016419049352407455, + 0.007811937481164932, + 0.00535172363743186, + 0.006946625653654337, + 0.0037804560270160437, + 0.010109602473676205, + 0.008580471388995647, + 0.004537608008831739, + 0.0028472119010984898, + 0.003702135756611824, + 0.0023867464624345303, + 0.003964777570217848, + 0.0032652171794325113, + 0.004197951406240463, + 0.0036921652499586344, + 0.006251344922930002, + 0.0035425343085080385, + 0.006433612667024136, + 0.008971106261014938, + 0.006638339255005121, + 0.00957638118416071, + 0.015998046845197678, + 0.010205782018601894, + 0.021296368911862373, + 0.011007101275026798, + 0.009111743420362473, + 0.015782175585627556, + 0.005622074007987976, + 0.005745119880884886, + 0.020243721082806587, + 0.01573823019862175, + 0.007917961105704308, + 0.007448152638971806, + 0.007609134539961815, + 0.013516951352357864, + 0.014154708944261074, + 0.009552372619509697, + 0.012960090301930904, + 0.00749570969492197, + 0.010252634063363075, + 0.008290818892419338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.02983754687011242 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07133467495441437, + 0.028825411573052406, + 0.03209085389971733, + 0.027803758159279823, + 0.10569426417350769, + 0.20894688367843628, + 0.05690218508243561, + 0.32190418243408203, + 0.10056044161319733, + 0.04593733325600624 + ], + "credit_map": [ + 0.005740929860621691, + 0.009829210117459297, + 0.0071176388300955296, + 0.012238132767379284, + 0.012642833404242992, + 0.01150916051119566, + 0.0069419750943779945, + 0.005314790178090334, + 0.004944474436342716, + 0.004484767559915781, + 0.003882126184180379, + 0.003548165550455451, + 0.00379434903152287, + 0.004023867659270763, + 0.0020855474285781384, + 0.0020621102303266525, + 0.003418942913413048, + 0.00398681778460741, + 0.005973713006824255, + 0.003456228179857135, + 0.003309705061838031, + 0.004662304185330868, + 0.004691879730671644, + 0.0025912607088685036, + 0.0027941677253693342, + 0.004196865018457174, + 0.0040983217768371105, + 0.0021678656339645386, + 0.0022983825765550137, + 0.0036331897135823965, + 0.0049515594728291035, + 0.003663404146209359, + 0.014907272532582283, + 0.019131334498524666, + 0.010891263373196125, + 0.010883854702115059, + 0.017923973500728607, + 0.01552200224250555, + 0.008984302170574665, + 0.00745026022195816, + 0.02875804901123047, + 0.027761947363615036, + 0.017781134694814682, + 0.022320756688714027, + 0.038213882595300674, + 0.03175394609570503, + 0.030177323147654533, + 0.012179825454950333, + 0.008740168996155262, + 0.005884263664484024, + 0.00633030291646719, + 0.00588490953668952, + 0.006003655958920717, + 0.008336319588124752, + 0.009120439179241657, + 0.006602122448384762, + 0.021556973457336426, + 0.035045698285102844, + 0.028756624087691307, + 0.03307764232158661, + 0.06434831768274307, + 0.03888235613703728, + 0.06321835517883301, + 0.037018194794654846, + 0.009928540326654911, + 0.019207289442420006, + 0.005564490798860788, + 0.0070847393944859505, + 0.022125042974948883, + 0.01875845156610012, + 0.009070186875760555, + 0.008821693249046803, + 0.006127688568085432, + 0.007397156208753586, + 0.006023386027663946, + 0.0048088934272527695, + 0.007745340466499329, + 0.005493114236742258, + 0.0044310251250863075, + 0.003910725936293602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "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 scene set in the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281339168548584 + ], + "opposite_logprob_tokens": [ + -1.406339168548584 + ], + "logprob_reward": -0.281339168548584, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281339168548584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1653410643339157, + 0.04957593232393265, + 0.030174685642123222, + 0.04301564395427704, + 0.11875320225954056, + 0.18016654253005981, + 0.03259918838739395, + 0.20860865712165833, + 0.11578914523124695, + 0.05597582459449768 + ], + "credit_map": [ + 0.014254039153456688, + 0.021127576008439064, + 0.018506936728954315, + 0.0106417341157794, + 0.04595539718866348, + 0.03034023381769657, + 0.01190098188817501, + 0.012614179402589798, + 0.007256703916937113, + 0.0072336927987635136, + 0.00555062061175704, + 0.0075159696862101555, + 0.006024899892508984, + 0.006946929730474949, + 0.004887791350483894, + 0.004159328993409872, + 0.0038300121668726206, + 0.003942845389246941, + 0.004864053800702095, + 0.002830647863447666, + 0.0028754672966897488, + 0.004734806250780821, + 0.00447502639144659, + 0.0026218295097351074, + 0.004819630645215511, + 0.006796868052333593, + 0.005108255427330732, + 0.003953527193516493, + 0.004866819363087416, + 0.005288318730890751, + 0.007340799085795879, + 0.004841430112719536, + 0.013299349695444107, + 0.02687000297009945, + 0.011836407706141472, + 0.010145273990929127, + 0.01882854662835598, + 0.01857471466064453, + 0.010686712339520454, + 0.008512211963534355, + 0.02717110514640808, + 0.020724434405565262, + 0.018630005419254303, + 0.015807978808879852, + 0.04314568266272545, + 0.03047524206340313, + 0.015433569438755512, + 0.008778546005487442, + 0.0036904276348650455, + 0.0038990976754575968, + 0.003565967781469226, + 0.0033461228013038635, + 0.006043174769729376, + 0.0044967434369027615, + 0.004815846215933561, + 0.002741810865700245, + 0.015289586037397385, + 0.02268213778734207, + 0.017105478793382645, + 0.021836193278431892, + 0.03805279731750488, + 0.02516988292336464, + 0.04344567656517029, + 0.025026939809322357, + 0.013064809143543243, + 0.021703200414776802, + 0.008609269745647907, + 0.007418838795274496, + 0.026052450761198997, + 0.01904177851974964, + 0.010659958235919476, + 0.009238858707249165, + 0.00549220060929656, + 0.007828181609511375, + 0.010408117435872555, + 0.006821817718446255, + 0.007829553447663784, + 0.005200383253395557, + 0.0069738938473165035, + 0.0054216827265918255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.281339168548584 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292989552021027 + ], + "opposite_logprob_tokens": [ + -2.2379298210144043 + ], + "logprob_reward": -0.11292989552021027, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292989552021027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3489893674850464, + 0.025833122432231903, + 0.09387937188148499, + 0.07046614587306976, + 0.05328759178519249, + 0.09322406351566315, + 0.01825876347720623, + 0.13623282313346863, + 0.0797111839056015, + 0.08011753112077713 + ], + "credit_map": [ + 0.03517655283212662, + 0.04178372398018837, + 0.03903313726186752, + 0.02454567514359951, + 0.06579294055700302, + 0.0518803708255291, + 0.04819536954164505, + 0.04258161783218384, + 0.003542141290381551, + 0.0045533087104558945, + 0.00297738378867507, + 0.00280134123750031, + 0.003844185033813119, + 0.0035721580497920513, + 0.002218109555542469, + 0.002324494766071439, + 0.012599163688719273, + 0.010583514347672462, + 0.015646900981664658, + 0.007628711871802807, + 0.008196927607059479, + 0.013605362735688686, + 0.016959795728325844, + 0.008658992126584053, + 0.0077382843010127544, + 0.009511713869869709, + 0.00846847239881754, + 0.005011274944990873, + 0.006421416066586971, + 0.008257853798568249, + 0.016734929755330086, + 0.00832220446318388, + 0.006424628663808107, + 0.011415170505642891, + 0.004584022797644138, + 0.004615353886038065, + 0.00816008634865284, + 0.008619554340839386, + 0.005804593674838543, + 0.003664184594526887, + 0.013842437416315079, + 0.011615769937634468, + 0.007629138417541981, + 0.009014690294861794, + 0.022037262097001076, + 0.014656830579042435, + 0.009272686205804348, + 0.0051552606746554375, + 0.0018503523897379637, + 0.0015132828848436475, + 0.0021488491911441088, + 0.0020795129239559174, + 0.0022069544065743685, + 0.0026106827426701784, + 0.0038205126766115427, + 0.0020286180078983307, + 0.018703563138842583, + 0.026777377352118492, + 0.012703577056527138, + 0.009472422301769257, + 0.022403942421078682, + 0.02422238141298294, + 0.013160288333892822, + 0.008789274841547012, + 0.011774559505283833, + 0.019102798774838448, + 0.006413954310119152, + 0.006472440902143717, + 0.009044740349054337, + 0.006947965361177921, + 0.008228376507759094, + 0.011726348660886288, + 0.009835821576416492, + 0.017609531059861183, + 0.007121006026864052, + 0.0064465682953596115, + 0.016504893079400063, + 0.010697142221033573, + 0.0057847024872899055, + 0.0061178794130682945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.11292989552021027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a level angle, captured the historic moment in extreme slow motion. In the frame, a complex lunar lander, clad in gold and silver reflective foil, descended silently and majestically toward the camera. In the airless vacuum of space, the sole source of sunlight cast a sharp light, blinding one side of the lander's metal frame while plunging the other into an abyss of pure darkness. As the lander's circular footpads neared touchdown, its invisible thrusters ejected the fine gray lunar dust from beneath in graceful radial arcs, each particle's trajectory clearly visible as it floated briefly in the weightless environment before cascading back down. Finally, the lander gently and steadily touched the lunar surface, stirring up the final puff of dust that slowly dissipated, and the world fell into absolute silence. Then, an astronaut, clad in a bulky but functional white spacesuit, his helmet covered with gold visors, carefully descended from the lander's ramp. Every movement is rendered with an air of exceptional lightness and poetry by the low gravity and slow motion. When his massive, deeply grooved boot finally touches the soft lunar surface, its sole sinks gently into the powdery lunar soil, leaving a clear, perfect, and eternal footprint. The camera slowly pulls out, framing the astronaut's tiny figure and the lander behind him, revealing the vast, desolate lunar landscape behind them, studded with craters and rocks, stretching to a curved horizon where it meets the inky blackness of the universe. This creates an epic atmosphere of grandeur, loneliness, and human courage.", + "seed": 24618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_3553_seed_24618_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": -0.07546082772314548, + "vlm_logprob": -0.07546082772314548, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07546082772314548, + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04215693473815918 + ], + "opposite_logprob_tokens": [ + -5.542157173156738 + ], + "logprob_reward": -0.04215693473815918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.04215693473815918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08631781488656998, + 0.042104605585336685, + 0.049505654722452164, + 0.13551770150661469, + 0.15289713442325592, + 0.02709375135600567, + 0.21162837743759155, + 0.1244802251458168, + 0.08002764731645584, + 0.09042704850435257 + ], + "credit_map": [ + 0.006366507150232792, + 0.01079055666923523, + 0.010807556100189686, + 0.014818636700510979, + 0.014800777658820152, + 0.013272645883262157, + 0.009868418797850609, + 0.005592720117419958, + 0.003838818985968828, + 0.0033644959330558777, + 0.005218415521085262, + 0.0037706007715314627, + 0.009994059801101685, + 0.007750633172690868, + 0.005719099659472704, + 0.0024484803434461355, + 0.004282649140805006, + 0.006889731157571077, + 0.008647257462143898, + 0.007740915287286043, + 0.006162042263895273, + 0.005118382629007101, + 0.0056149400770664215, + 0.005049733445048332, + 0.006416034884750843, + 0.008335874415934086, + 0.022146398201584816, + 0.04352109506726265, + 0.007969537749886513, + 0.015353331342339516, + 0.022535689175128937, + 0.0092397416010499, + 0.008373316377401352, + 0.005255557131022215, + 0.05683965981006622, + 0.05602208524942398, + 0.003308901097625494, + 0.005459133069962263, + 0.011256799101829529, + 0.00638167466968298, + 0.003557284362614155, + 0.004713278263807297, + 0.004805890843272209, + 0.0041961087845265865, + 0.0019808760844171047, + 0.0014999309787526727, + 0.0037282882258296013, + 0.0026120946276932955, + 0.015971319749951363, + 0.015676744282245636, + 0.06753091514110565, + 0.020007621496915817, + 0.010445979423820972, + 0.012409630231559277, + 0.02678193524479866, + 0.04280423745512962, + 0.011639533564448357, + 0.012344422750175, + 0.024856872856616974, + 0.014093067497015, + 0.01565750502049923, + 0.011160893365740776, + 0.01789395697414875, + 0.016833966597914696, + 0.00785161554813385, + 0.01067778654396534, + 0.012194792740046978, + 0.00873641762882471, + 0.011131851002573967, + 0.006122905295342207, + 0.01234204787760973, + 0.010970231145620346, + 0.011674371547996998, + 0.011860323138535023, + 0.013842311687767506, + 0.010826601646840572, + 0.01216635201126337, + 0.008809076622128487, + 0.011113578453660011, + 0.010134438052773476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.04215693473815918 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026372436434030533 + ], + "opposite_logprob_tokens": [ + -3.6513724327087402 + ], + "logprob_reward": -0.026372436434030533, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026372436434030533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08197720348834991, + 0.033293675631284714, + 0.06356360763311386, + 0.334745317697525, + 0.18353042006492615, + 0.03906184807419777, + 0.045824818313121796, + 0.08463256806135178, + 0.0922418087720871, + 0.041128724813461304 + ], + "credit_map": [ + 0.009785170666873455, + 0.013558482751250267, + 0.011852234601974487, + 0.009960499592125416, + 0.011483547277748585, + 0.010526482947170734, + 0.0085276635363698, + 0.006283125374466181, + 0.002557563129812479, + 0.003355913795530796, + 0.0037642596289515495, + 0.004655070137232542, + 0.0070646293461322784, + 0.005434628110378981, + 0.0041193184442818165, + 0.002342294668778777, + 0.004436031449586153, + 0.0059126331470906734, + 0.009884431958198547, + 0.01564173400402069, + 0.004764568991959095, + 0.007281286176294088, + 0.0068586780689656734, + 0.00878424383699894, + 0.01376386359333992, + 0.016412151977419853, + 0.05580800026655197, + 0.12715697288513184, + 0.01594436727464199, + 0.049135591834783554, + 0.036418430507183075, + 0.020105933770537376, + 0.010690066032111645, + 0.007702124770730734, + 0.07346244901418686, + 0.06134571135044098, + 0.004165951628237963, + 0.0061019123531877995, + 0.014303450472652912, + 0.00575875723734498, + 0.004602550994604826, + 0.004129740875214338, + 0.006712185684591532, + 0.006138740107417107, + 0.0030571776442229748, + 0.002669753273949027, + 0.006188681349158287, + 0.005563017912209034, + 0.003631697967648506, + 0.002628170419484377, + 0.01347840391099453, + 0.005187004338949919, + 0.0022273214999586344, + 0.0027353912591934204, + 0.006240542978048325, + 0.009696288965642452, + 0.009041588753461838, + 0.009764091111719608, + 0.012467076070606709, + 0.015476190485060215, + 0.010270174592733383, + 0.006205981597304344, + 0.010869304649531841, + 0.010538159869611263, + 0.00794930849224329, + 0.014786972664296627, + 0.015744635835289955, + 0.012370769865810871, + 0.008069276809692383, + 0.007412065286189318, + 0.01366361416876316, + 0.012245165184140205, + 0.005084698088467121, + 0.008142896927893162, + 0.006092080846428871, + 0.004201468545943499, + 0.0033295287285000086, + 0.004152523819357157, + 0.0052943844348192215, + 0.004831141326576471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.026372436434030533 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07162714749574661, + 0.02622915804386139, + 0.15988053381443024, + 0.09730727970600128, + 0.1443626880645752, + 0.019332775846123695, + 0.1383943408727646, + 0.11351989954710007, + 0.12166449427604675, + 0.10768163204193115 + ], + "credit_map": [ + 0.004962497390806675, + 0.00809633918106556, + 0.007973911240696907, + 0.007136851083487272, + 0.011963006108999252, + 0.010104354470968246, + 0.011229357682168484, + 0.01016082800924778, + 0.0017963520949706435, + 0.0035310015082359314, + 0.00401332788169384, + 0.0034877301659435034, + 0.00468906294554472, + 0.0040508597157895565, + 0.002786438213661313, + 0.0018743849359452724, + 0.012236097827553749, + 0.02246377058327198, + 0.025875123217701912, + 0.024549083784222603, + 0.023479662835597992, + 0.017983760684728622, + 0.018201660364866257, + 0.015091384761035442, + 0.004084530286490917, + 0.005293700378388166, + 0.0160460714250803, + 0.034740205854177475, + 0.005509523209184408, + 0.013893825933337212, + 0.010834846645593643, + 0.0069045741111040115, + 0.008812496438622475, + 0.007511647883802652, + 0.055906087160110474, + 0.04582023620605469, + 0.004168050363659859, + 0.0039052932988852262, + 0.013118766248226166, + 0.005120111163705587, + 0.0017036722274497151, + 0.001976015977561474, + 0.003468873677775264, + 0.00377947511151433, + 0.0019157362403348088, + 0.0012467870255932212, + 0.003124380484223366, + 0.0021178347524255514, + 0.011962823569774628, + 0.012312476523220539, + 0.03545227274298668, + 0.015578551217913628, + 0.007678421214222908, + 0.008859132416546345, + 0.01793290115892887, + 0.02861776389181614, + 0.014366690069437027, + 0.01450904831290245, + 0.018414506688714027, + 0.013554330915212631, + 0.014497659169137478, + 0.009626483544707298, + 0.014817333780229092, + 0.013733846135437489, + 0.013517399318516254, + 0.02491457387804985, + 0.02003062516450882, + 0.013994311913847923, + 0.010177790187299252, + 0.009664294309914112, + 0.013956739567220211, + 0.015408762730658054, + 0.011673256754875183, + 0.016584446653723717, + 0.01343152578920126, + 0.012721890583634377, + 0.013783828355371952, + 0.009332315064966679, + 0.01625589095056057, + 0.013898474164307117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "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 scene set in the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156078040599823 + ], + "opposite_logprob_tokens": [ + -1.7015607357025146 + ], + "logprob_reward": -0.20156078040599823, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156078040599823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10399402678012848, + 0.04526226967573166, + 0.16067837178707123, + 0.17961618304252625, + 0.17343075573444366, + 0.019995292648673058, + 0.058310601860284805, + 0.06661375612020493, + 0.13878369331359863, + 0.05331505089998245 + ], + "credit_map": [ + 0.006138088181614876, + 0.008840366266667843, + 0.011235935613512993, + 0.009575044736266136, + 0.028854571282863617, + 0.01934872753918171, + 0.01172288041561842, + 0.008278406225144863, + 0.002521171700209379, + 0.003604724770411849, + 0.0050669764168560505, + 0.006111035123467445, + 0.011084847152233124, + 0.007999858818948269, + 0.005995861254632473, + 0.0028777923434972763, + 0.013545638881623745, + 0.020834602415561676, + 0.027793942019343376, + 0.026301654055714607, + 0.023156674578785896, + 0.018963545560836792, + 0.017368542030453682, + 0.012713775970041752, + 0.007187015376985073, + 0.00920168124139309, + 0.02858855575323105, + 0.07083380222320557, + 0.007890067994594574, + 0.02319372445344925, + 0.022439980879426003, + 0.010281362570822239, + 0.00862623006105423, + 0.006942277308553457, + 0.07146894931793213, + 0.05792901664972305, + 0.003988169599324465, + 0.0044381762854754925, + 0.014570987783372402, + 0.005466939881443977, + 0.002367001259699464, + 0.0021611307747662067, + 0.002944166073575616, + 0.0033070952631533146, + 0.0020837620832026005, + 0.001230881316587329, + 0.0032805867958813906, + 0.002620670013129711, + 0.005623471923172474, + 0.005645326804369688, + 0.013270760886371136, + 0.006753967609256506, + 0.004773436579853296, + 0.003195288358256221, + 0.010004354640841484, + 0.00904399435967207, + 0.007881826721131802, + 0.006703174207359552, + 0.009470642544329166, + 0.007440090645104647, + 0.010154491290450096, + 0.007717274595052004, + 0.009898942895233631, + 0.0073473090305924416, + 0.014181096106767654, + 0.012474394403398037, + 0.015821706503629684, + 0.013261680491268635, + 0.024730337783694267, + 0.00992631260305643, + 0.028997277840971947, + 0.019390881061553955, + 0.005743866786360741, + 0.007264604791998863, + 0.0070785400457680225, + 0.0056823622435331345, + 0.006465195678174496, + 0.004631415940821171, + 0.008832426741719246, + 0.007616640534251928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.20156078040599823 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902223408222198 + ], + "opposite_logprob_tokens": [ + -2.464022159576416 + ], + "logprob_reward": -0.08902223408222198, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902223408222198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04294221103191376, + 0.03012729622423649, + 0.02346161939203739, + 0.10221648216247559, + 0.023527782410383224, + 0.01810624450445175, + 0.20958198606967926, + 0.13278929889202118, + 0.09560611844062805, + 0.32164090871810913 + ], + "credit_map": [ + 0.0029471097514033318, + 0.003408175427466631, + 0.005358211230486631, + 0.005738535430282354, + 0.008030636236071587, + 0.006417764350771904, + 0.005225983448326588, + 0.0058157918974757195, + 0.0024714882019907236, + 0.0024153389967978, + 0.0034298731479793787, + 0.004895682446658611, + 0.006612993311136961, + 0.0050248377956449986, + 0.0029135304503142834, + 0.0023635542020201683, + 0.002396909985691309, + 0.0028855528216809034, + 0.0038262782618403435, + 0.0036573661491274834, + 0.003880901727825403, + 0.002870178548619151, + 0.0019314141245558858, + 0.002013016492128372, + 0.0044140201061964035, + 0.005323960445821285, + 0.01649627462029457, + 0.03203350678086281, + 0.006584127899259329, + 0.010138826444745064, + 0.021272940561175346, + 0.005952821578830481, + 0.0029359047766774893, + 0.0022658947855234146, + 0.0042248680256307125, + 0.0038988436572253704, + 0.0015935010742396116, + 0.0023230432998389006, + 0.0036411972250789404, + 0.002644529100507498, + 0.002151601016521454, + 0.0019514559535309672, + 0.0029654938261955976, + 0.0030806069262325764, + 0.0015731266466900706, + 0.001002504606731236, + 0.003271637484431267, + 0.002109817462041974, + 0.016342245042324066, + 0.01443502027541399, + 0.06618358939886093, + 0.01920432224869728, + 0.008768972009420395, + 0.012289431877434254, + 0.02865191549062729, + 0.043706510215997696, + 0.014508543536067009, + 0.01057516224682331, + 0.025558672845363617, + 0.01930798403918743, + 0.01078723557293415, + 0.015922188758850098, + 0.016285493969917297, + 0.019844021648168564, + 0.007724662311375141, + 0.009560594335198402, + 0.016026001423597336, + 0.011890185996890068, + 0.009313646703958511, + 0.005901001393795013, + 0.02048349566757679, + 0.014706525020301342, + 0.038386158645153046, + 0.054514478892087936, + 0.041732095181941986, + 0.03141559660434723, + 0.033961910754442215, + 0.02860676683485508, + 0.05169708654284477, + 0.041326817125082016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.08902223408222198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from a level angle, captured the historic moment in extreme slow motion. In the frame, a complex lunar lander, clad in gold and silver reflective foil, descended silently and majestically toward the camera. In the airless vacuum of space, the sole source of sunlight cast a sharp light, blinding one side of the lander's metal frame while plunging the other into an abyss of pure darkness. As the lander's circular footpads neared touchdown, its invisible thrusters ejected the fine gray lunar dust from beneath in graceful radial arcs, each particle's trajectory clearly visible as it floated briefly in the weightless environment before cascading back down. Finally, the lander gently and steadily touched the lunar surface, stirring up the final puff of dust that slowly dissipated, and the world fell into absolute silence. Then, an astronaut, clad in a bulky but functional white spacesuit, his helmet covered with gold visors, carefully descended from the lander's ramp. Every movement is rendered with an air of exceptional lightness and poetry by the low gravity and slow motion. When his massive, deeply grooved boot finally touches the soft lunar surface, its sole sinks gently into the powdery lunar soil, leaving a clear, perfect, and eternal footprint. The camera slowly pulls out, framing the astronaut's tiny figure and the lander behind him, revealing the vast, desolate lunar landscape behind them, studded with craters and rocks, stretching to a curved horizon where it meets the inky blackness of the universe. This creates an epic atmosphere of grandeur, loneliness, and human courage.", + "seed": 24619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_5/index_3553_seed_24619_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": -0.04676107047125697, + "vlm_logprob": -0.04676107047125697, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04676107047125697, + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06343971937894821 + ], + "opposite_logprob_tokens": [ + -4.813439846038818 + ], + "logprob_reward": -0.06343971937894821, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.06343971937894821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15533943474292755, + 0.09768842160701752, + 0.0659683346748352, + 0.07515135407447815, + 0.09646885842084885, + 0.10245531052350998, + 0.09744221717119217, + 0.12132349610328674, + 0.05655728280544281, + 0.13160528242588043 + ], + "credit_map": [ + 0.02055964060127735, + 0.012284224852919579, + 0.024909986183047295, + 0.029603658244013786, + 0.02363801933825016, + 0.01081113237887621, + 0.013738296926021576, + 0.019794471561908722, + 0.01205655187368393, + 0.00796812679618597, + 0.017643142491579056, + 0.01127699762582779, + 0.01114918664097786, + 0.009861728176474571, + 0.013282429426908493, + 0.01445025485008955, + 0.008786355145275593, + 0.006524398922920227, + 0.006370989140123129, + 0.007921459153294563, + 0.006483947392553091, + 0.007973105646669865, + 0.011121541261672974, + 0.010786538943648338, + 0.006562706083059311, + 0.007113128434866667, + 0.0092748012393713, + 0.010695231147110462, + 0.007770259864628315, + 0.00812466535717249, + 0.011320157907903194, + 0.0142904007807374, + 0.007088820915669203, + 0.006674543954432011, + 0.014380785636603832, + 0.013923895545303822, + 0.008932163007557392, + 0.009289470501244068, + 0.015784624963998795, + 0.02039455994963646, + 0.010040687397122383, + 0.009512634947896004, + 0.01648012176156044, + 0.014267700724303722, + 0.007946951314806938, + 0.008499964140355587, + 0.016957100480794907, + 0.018750153481960297, + 0.008470206521451473, + 0.0123281953856349, + 0.014975447207689285, + 0.012754094786942005, + 0.009943142533302307, + 0.011318594217300415, + 0.014018411748111248, + 0.013634122908115387, + 0.009232197888195515, + 0.023407746106386185, + 0.015863332897424698, + 0.010207467712461948, + 0.02808765508234501, + 0.018113603815436363, + 0.010272887535393238, + 0.0061386036686599255, + 0.007331511005759239, + 0.011273273266851902, + 0.007703960407525301, + 0.0076691447757184505, + 0.005534749012440443, + 0.0071669346652925014, + 0.0046064057387411594, + 0.0052713071927428246, + 0.014402895234525204, + 0.02807261422276497, + 0.017289411276578903, + 0.025293920189142227, + 0.012302720919251442, + 0.013049859553575516, + 0.010876142419874668, + 0.010317720472812653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.06343971937894821 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007659589406102896 + ], + "opposite_logprob_tokens": [ + -4.882659435272217 + ], + "logprob_reward": -0.007659589406102896, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007659589406102896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16405712068080902, + 0.1322527974843979, + 0.10204263776540756, + 0.12248383462429047, + 0.08773506432771683, + 0.09071177989244461, + 0.09401948004961014, + 0.04158126562833786, + 0.042519569396972656, + 0.12259644269943237 + ], + "credit_map": [ + 0.022576889023184776, + 0.01416572742164135, + 0.02220413088798523, + 0.024958660826086998, + 0.019687432795763016, + 0.014993474818766117, + 0.021206291392445564, + 0.024264518171548843, + 0.023407503962516785, + 0.014409330673515797, + 0.013907033018767834, + 0.014926964417099953, + 0.02643839828670025, + 0.016077818349003792, + 0.014674341306090355, + 0.008411397226154804, + 0.01132568996399641, + 0.008618912659585476, + 0.014806567691266537, + 0.016492120921611786, + 0.01049741543829441, + 0.009067581966519356, + 0.014278239570558071, + 0.016956105828285217, + 0.0113450912758708, + 0.012138067744672298, + 0.024156812578439713, + 0.02192562073469162, + 0.011058955453336239, + 0.013984736055135727, + 0.015394961461424828, + 0.012479581870138645, + 0.005670613143593073, + 0.0065280478447675705, + 0.012688842602074146, + 0.011288408190011978, + 0.008952277712523937, + 0.00793676357716322, + 0.015535511076450348, + 0.01913459599018097, + 0.011372753418982029, + 0.008333600126206875, + 0.013021468184888363, + 0.00994187407195568, + 0.009090700186789036, + 0.00788672361522913, + 0.01467615645378828, + 0.016388503834605217, + 0.011995089240372181, + 0.008960836566984653, + 0.01448843814432621, + 0.012566697783768177, + 0.00995370652526617, + 0.010440467856824398, + 0.013998596929013729, + 0.011615647003054619, + 0.003523855470120907, + 0.00742101576179266, + 0.005685437936335802, + 0.003814052091911435, + 0.007905482314527035, + 0.005918082781136036, + 0.0037912162952125072, + 0.003522126702591777, + 0.0038320450112223625, + 0.008909163996577263, + 0.005905107595026493, + 0.007063583470880985, + 0.005705616436898708, + 0.004739147145301104, + 0.00395775493234396, + 0.0024071524385362864, + 0.013220367953181267, + 0.014368213713169098, + 0.013443434610962868, + 0.01678546331822872, + 0.022732427343726158, + 0.017096227034926414, + 0.011991274543106556, + 0.01295903418213129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.007659589406102896 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.15030717849731445, + 0.32352185249328613, + 0.09830231964588165, + 0.0984368845820427, + 0.07324875146150589, + 0.02990441583096981, + 0.06812632828950882, + 0.030446279793977737, + 0.08685989677906036, + 0.040846049785614014 + ], + "credit_map": [ + 0.016907064244151115, + 0.011331799440085888, + 0.02412267029285431, + 0.034354060888290405, + 0.020146667957305908, + 0.014124349690973759, + 0.012966626323759556, + 0.016353929415345192, + 0.04311656951904297, + 0.021059280261397362, + 0.054660048335790634, + 0.045884352177381516, + 0.02367871068418026, + 0.021626748144626617, + 0.08033852279186249, + 0.03315761312842369, + 0.013901923783123493, + 0.006271866150200367, + 0.017332175746560097, + 0.013496607542037964, + 0.010627850890159607, + 0.01001801062375307, + 0.014081185683608055, + 0.012572700157761574, + 0.008432477712631226, + 0.007909756153821945, + 0.021064748987555504, + 0.019545918330550194, + 0.007654642686247826, + 0.009825666435062885, + 0.012945184484124184, + 0.011058487929403782, + 0.0051190052181482315, + 0.004223335068672895, + 0.010833652690052986, + 0.010014916770160198, + 0.006016923580318689, + 0.00499438401311636, + 0.016317253932356834, + 0.015729280188679695, + 0.00334380310960114, + 0.0033403353299945593, + 0.0053194682113826275, + 0.004463326185941696, + 0.0025600322987884283, + 0.0025623636320233345, + 0.004717577248811722, + 0.003597511211410165, + 0.008413100615143776, + 0.00572236767038703, + 0.010775845497846603, + 0.008111177012324333, + 0.006730230059474707, + 0.01035983581095934, + 0.010683190077543259, + 0.007330584805458784, + 0.0029161248821765184, + 0.004382098093628883, + 0.0037449810188263655, + 0.0029060726519674063, + 0.005362520460039377, + 0.0033540036529302597, + 0.0044151488691568375, + 0.003365331795066595, + 0.007836072705686092, + 0.01758970506489277, + 0.01392932329326868, + 0.01361396349966526, + 0.011114390566945076, + 0.009831493720412254, + 0.006381535902619362, + 0.006563406903296709, + 0.004651141352951527, + 0.006901714019477367, + 0.005034964065998793, + 0.006622537970542908, + 0.006773095112293959, + 0.004377242177724838, + 0.0037562462966889143, + 0.00272911018691957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01607704907655716 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698562532663345 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.033698562532663345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698562532663345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14394567906856537, + 0.04680224880576134, + 0.09089908748865128, + 0.06062730774283409, + 0.07848068326711655, + 0.06317336857318878, + 0.11097841709852219, + 0.08546049892902374, + 0.1499636173248291, + 0.16966912150382996 + ], + "credit_map": [ + 0.017887193709611893, + 0.010564667172729969, + 0.01939784176647663, + 0.021797874942421913, + 0.02144448459148407, + 0.014105335809290409, + 0.024399779736995697, + 0.014348506927490234, + 0.0070966738276183605, + 0.004633639007806778, + 0.009810389950871468, + 0.006352209020406008, + 0.004716023802757263, + 0.0047113592736423016, + 0.006088788155466318, + 0.0033931657671928406, + 0.012896631844341755, + 0.0075394222512841225, + 0.015933392569422722, + 0.009908567182719707, + 0.010786784812808037, + 0.011261430568993092, + 0.013293826952576637, + 0.009279035963118076, + 0.004575673490762711, + 0.004225083161145449, + 0.017007509246468544, + 0.008572268299758434, + 0.004912062548100948, + 0.005394009407609701, + 0.010833323001861572, + 0.005107379984110594, + 0.005123374983668327, + 0.00510908430442214, + 0.01122979260981083, + 0.009937276132404804, + 0.006797158159315586, + 0.006765749305486679, + 0.018846407532691956, + 0.014671838842332363, + 0.0076977647840976715, + 0.005438291933387518, + 0.009018371812999249, + 0.009293999522924423, + 0.0067558917216956615, + 0.007634240668267012, + 0.007936405017971992, + 0.009398397989571095, + 0.012382971122860909, + 0.012261536903679371, + 0.015528025105595589, + 0.010636678896844387, + 0.018391985446214676, + 0.018950045108795166, + 0.011675865389406681, + 0.011151306331157684, + 0.007570568937808275, + 0.01844540424644947, + 0.012862141244113445, + 0.01000453531742096, + 0.009685434401035309, + 0.009911123663187027, + 0.00969682540744543, + 0.0072844745591282845, + 0.01512902695685625, + 0.03341533616185188, + 0.023547882214188576, + 0.023872409015893936, + 0.015744304284453392, + 0.01607058197259903, + 0.009874486364424229, + 0.012309575453400612, + 0.01682894676923752, + 0.038325853645801544, + 0.022715959697961807, + 0.03385869413614273, + 0.014820876531302929, + 0.015450665727257729, + 0.015240889973938465, + 0.012427247129380703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.033698562532663345 + }, + { + "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 the airless vacuum of space on the lunar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a complex lunar lander clad in gold and silver reflective foil visible descending toward the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lander eject fine gray lunar dust in graceful radial arcs as it nears touchdown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an astronaut wearing a white spacesuit with a gold visor helmet visible descending from the lander's ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pull out to reveal the vast lunar landscape with craters and a curved horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293043196201324 + ], + "opposite_logprob_tokens": [ + -2.2379305362701416 + ], + "logprob_reward": -0.11293043196201324, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293043196201324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13834883272647858, + 0.06006820499897003, + 0.06152034550905228, + 0.04336303472518921, + 0.055323511362075806, + 0.09541508555412292, + 0.2074974775314331, + 0.15457749366760254, + 0.13954584300518036, + 0.04434007778763771 + ], + "credit_map": [ + 0.01689901202917099, + 0.009827641770243645, + 0.02021176740527153, + 0.021276013925671577, + 0.01077538263052702, + 0.010325652547180653, + 0.028865303844213486, + 0.020168060436844826, + 0.009044745936989784, + 0.005860843695700169, + 0.00865423958748579, + 0.004870711825788021, + 0.010739993304014206, + 0.007783731445670128, + 0.007952125743031502, + 0.005161810200661421, + 0.008993868716061115, + 0.004995635244995356, + 0.006163702346384525, + 0.0078618498519063, + 0.007990016601979733, + 0.008779426105320454, + 0.00997893139719963, + 0.0067569115199148655, + 0.0035941051319241524, + 0.002652654657140374, + 0.008345578797161579, + 0.008355479687452316, + 0.0036795008927583694, + 0.004457381553947926, + 0.006936350371688604, + 0.005341982934623957, + 0.004188189283013344, + 0.003808302339166403, + 0.008324330672621727, + 0.00670476583763957, + 0.0045145247131586075, + 0.004922100342810154, + 0.011561502702534199, + 0.01129979733377695, + 0.011887245811522007, + 0.00782511942088604, + 0.017775295302271843, + 0.01138248760253191, + 0.008877627551555634, + 0.009751927107572556, + 0.016871457919478416, + 0.01104392297565937, + 0.026821963489055634, + 0.021397879347205162, + 0.03428422287106514, + 0.023927921429276466, + 0.014775150455534458, + 0.027655886486172676, + 0.03744566813111305, + 0.02118876948952675, + 0.01392732746899128, + 0.024857070297002792, + 0.019365480169653893, + 0.01880274899303913, + 0.02535903826355934, + 0.01759936660528183, + 0.0182558074593544, + 0.01641065813601017, + 0.014107098802924156, + 0.02861495316028595, + 0.02247609570622444, + 0.021720144897699356, + 0.015615311451256275, + 0.015014270320534706, + 0.010905970819294453, + 0.011092009954154491, + 0.005859259981662035, + 0.009037503041327, + 0.005657382775098085, + 0.006498611532151699, + 0.004370551090687513, + 0.004747822415083647, + 0.004336578771471977, + 0.003832368878647685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.11293043196201324 + } + ], + "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/0000060/group_6/index_tensor([3324, 3324])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_tensor([3324, 3324])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b431a443624a1d56a3b0d0a4df24377cc37c36c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_tensor([3324, 3324])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a giant wild boar, as massive as a calf, charging forward with unstoppable force deep within an ancient, dense forest. Towering trees intertwine their branches, nearly obscuring the sky. Only a few faint, grayish-white rays of light penetrate through the gaps in the canopy, casting shifting, dappled shadows on the ground. The boar's mane, coarse, dark brown and black, is matted with damp earth and crushed plant debris. A pair of curved, yellowed tusks jut from its massive, lowered head, and its small, fierce eyes gleam with a wild glint. Each thump of its powerful hooves on the slippery, black earth studded with rotting leaves sends chunks of mud and years' worth of dead branches and leaves flying violently backward and to the sides, creating a chaotic, swirling cloud of debris visible in the dim light. It crushes small saplings and shrubs in its path. The main colors of the entire picture are deep brown and dark green, full of primitive, tense and chaotic dynamics, perfectly presenting the wild power of the beast rushing through the dense forest and being unstoppable.", + "seed": 6705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6705_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": -0.009582957345992326, + "vlm_logprob": -0.009582957345992326, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.009582957345992326, + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014741425402462482 + ], + "opposite_logprob_tokens": [ + -4.38974142074585 + ], + "logprob_reward": -0.014741425402462482, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014741425402462482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22749976813793182, + 0.028889372944831848, + 0.02252017892897129, + 0.10713927447795868, + 0.055898163467645645, + 0.06931557506322861, + 0.04000091552734375, + 0.04804542660713196, + 0.18721747398376465, + 0.2134738713502884 + ], + "credit_map": [ + 0.02906152792274952, + 0.009858530946075916, + 0.07406633347272873, + 0.01063618715852499, + 0.008803081698715687, + 0.06339140981435776, + 0.011133834719657898, + 0.020548881962895393, + 0.0032695718109607697, + 0.00312624080106616, + 0.004215986933559179, + 0.003695548279210925, + 0.003318725386634469, + 0.003063975367695093, + 0.0035744505003094673, + 0.004624873399734497, + 0.003274473361670971, + 0.0017215803964063525, + 0.003188914153724909, + 0.0023184558376669884, + 0.0033829996827989817, + 0.002574248006567359, + 0.003146940842270851, + 0.0029125667642802, + 0.008613244630396366, + 0.010141691192984581, + 0.013522356748580933, + 0.010922150686383247, + 0.019952552393078804, + 0.007997089996933937, + 0.02218707837164402, + 0.013803106732666492, + 0.003359232097864151, + 0.007285736035555601, + 0.007202783599495888, + 0.005492417141795158, + 0.009863210842013359, + 0.004417741671204567, + 0.010582096874713898, + 0.007694946136325598, + 0.0061653791926801205, + 0.007051172666251659, + 0.00946173258125782, + 0.006933853030204773, + 0.00937893521040678, + 0.006969624198973179, + 0.013610630296170712, + 0.00974425021559, + 0.0032609011977910995, + 0.0029416889883577824, + 0.00432482548058033, + 0.002981453202664852, + 0.006148250307887793, + 0.008341852575540543, + 0.004517276771366596, + 0.0074846684001386166, + 0.006022896151989698, + 0.004309321288019419, + 0.003711846424266696, + 0.005163040943443775, + 0.007255898788571358, + 0.006738130934536457, + 0.007096446584910154, + 0.007747843861579895, + 0.01731712743639946, + 0.024463944137096405, + 0.02135416492819786, + 0.028707044199109077, + 0.03442735597491264, + 0.015113999135792255, + 0.027066096663475037, + 0.018767734989523888, + 0.019259413704276085, + 0.028199944645166397, + 0.02655627951025963, + 0.03265129402279854, + 0.026599718257784843, + 0.028541458770632744, + 0.031726010143756866, + 0.019939756020903587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.014741425402462482 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607329398393631 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607329398393631, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607329398393631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15924610197544098, + 0.10136187821626663, + 0.02604253962635994, + 0.07661876082420349, + 0.07242333889007568, + 0.13067248463630676, + 0.038462791591882706, + 0.06344756484031677, + 0.13568046689033508, + 0.1960439831018448 + ], + "credit_map": [ + 0.017257776111364365, + 0.009324658662080765, + 0.03413503244519234, + 0.02037658356130123, + 0.009861822240054607, + 0.03152390196919441, + 0.01716044172644615, + 0.01960589922964573, + 0.00825203862041235, + 0.013459417037665844, + 0.014676922932267189, + 0.01161347609013319, + 0.009570763446390629, + 0.012415590696036816, + 0.012434088625013828, + 0.01893959753215313, + 0.0026291515678167343, + 0.0025560238864272833, + 0.003242805367335677, + 0.003392728278413415, + 0.002775820903480053, + 0.003711118595674634, + 0.0045961979776620865, + 0.0031386949121952057, + 0.005951831117272377, + 0.005386614240705967, + 0.009010188281536102, + 0.009563879109919071, + 0.011336666531860828, + 0.006101123057305813, + 0.018970131874084473, + 0.010298335924744606, + 0.005778547376394272, + 0.013019817881286144, + 0.013617352582514286, + 0.008680964820086956, + 0.00823319423943758, + 0.005498666316270828, + 0.008407854475080967, + 0.009186942130327225, + 0.011026890017092228, + 0.022460194304585457, + 0.017134876921772957, + 0.015613280236721039, + 0.012109469622373581, + 0.012424143962562084, + 0.022985020652413368, + 0.016918621957302094, + 0.005185800604522228, + 0.003953614272177219, + 0.004889821168035269, + 0.0026701013557612896, + 0.0037920454051345587, + 0.008291753940284252, + 0.004042139742523432, + 0.0056375195272266865, + 0.007652882020920515, + 0.006977397948503494, + 0.005631063599139452, + 0.005483055487275124, + 0.011274208314716816, + 0.008279195055365562, + 0.008753047324717045, + 0.009396721608936787, + 0.01633422262966633, + 0.016043569892644882, + 0.015368266962468624, + 0.012720960192382336, + 0.026122204959392548, + 0.013117768801748753, + 0.018426788970828056, + 0.017546702176332474, + 0.01781916618347168, + 0.0303292665630579, + 0.02457009255886078, + 0.026156656444072723, + 0.02601911872625351, + 0.02956376224756241, + 0.022953135892748833, + 0.018632804974913597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01607329398393631 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11787474155426025, + 0.04826311394572258, + 0.041518453508615494, + 0.08685297518968582, + 0.0716945081949234, + 0.1123744547367096, + 0.0773157924413681, + 0.08257672190666199, + 0.09766781330108643, + 0.26386135816574097 + ], + "credit_map": [ + 0.018585192039608955, + 0.008609859272837639, + 0.01960858888924122, + 0.01355794444680214, + 0.007276381831616163, + 0.014870385639369488, + 0.017859473824501038, + 0.017506923526525497, + 0.004593970719724894, + 0.0068148053251206875, + 0.005678380839526653, + 0.004856889136135578, + 0.005919173359870911, + 0.0056800213642418385, + 0.0065311104990541935, + 0.008188768289983273, + 0.004427414387464523, + 0.0037016835995018482, + 0.004358808510005474, + 0.004229282028973103, + 0.007537242025136948, + 0.00422378396615386, + 0.00728018069639802, + 0.005760065745562315, + 0.0065607307478785515, + 0.008722196333110332, + 0.009293615818023682, + 0.010418480262160301, + 0.01220397837460041, + 0.008763345889747143, + 0.020101625472307205, + 0.010789014399051666, + 0.006176694296300411, + 0.00868701096624136, + 0.0075612482614815235, + 0.0056441775523126125, + 0.01635567657649517, + 0.008463979698717594, + 0.007268864195793867, + 0.011536864563822746, + 0.007930447347462177, + 0.011475281789898872, + 0.01640361174941063, + 0.012957839295268059, + 0.014494536444544792, + 0.010927525348961353, + 0.022215040400624275, + 0.01597018539905548, + 0.004826230928301811, + 0.003750134725123644, + 0.003721589921042323, + 0.002935517579317093, + 0.008425494655966759, + 0.02743120677769184, + 0.005367483478039503, + 0.02085813879966736, + 0.006879066117107868, + 0.00700779166072607, + 0.004984414204955101, + 0.008726674132049084, + 0.016123797744512558, + 0.011829276569187641, + 0.015289043076336384, + 0.01173667423427105, + 0.005915631540119648, + 0.01548049133270979, + 0.010274628177285194, + 0.016484415158629417, + 0.01734582893550396, + 0.007449746131896973, + 0.014183846302330494, + 0.010533236898481846, + 0.019619833678007126, + 0.03388131782412529, + 0.025433670729398727, + 0.03532303497195244, + 0.03979383781552315, + 0.040185071527957916, + 0.038588106632232666, + 0.031036486849188805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.011074034497141838 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019369428046047688 + ], + "opposite_logprob_tokens": [ + -6.251936912536621 + ], + "logprob_reward": -0.0019369428046047688, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019369428046047688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30481868982315063, + 0.04635060578584671, + 0.03719846531748772, + 0.04934006184339523, + 0.044577423483133316, + 0.1857500523328781, + 0.0868210569024086, + 0.042935729026794434, + 0.0660126730799675, + 0.13619525730609894 + ], + "credit_map": [ + 0.03394851088523865, + 0.0198205653578043, + 0.0970519632101059, + 0.0169992633163929, + 0.012719542719423771, + 0.07578150928020477, + 0.017294658347964287, + 0.031202686950564384, + 0.005430485587567091, + 0.004212852101773024, + 0.005241523962467909, + 0.005425473675131798, + 0.004403872415423393, + 0.007814393378794193, + 0.005199007224291563, + 0.008623001165688038, + 0.003747117007151246, + 0.0035274894908070564, + 0.003670815844088793, + 0.003352157538756728, + 0.005751017015427351, + 0.0050718048587441444, + 0.006608664523810148, + 0.005469402298331261, + 0.004223939031362534, + 0.004877294413745403, + 0.006121508777141571, + 0.005139240995049477, + 0.008431664668023586, + 0.005118186119943857, + 0.008984326384961605, + 0.006443902384489775, + 0.003949107602238655, + 0.007003954611718655, + 0.00625587347894907, + 0.00491847051307559, + 0.005652179941534996, + 0.0056473566219210625, + 0.004750287160277367, + 0.0064001972787082195, + 0.026548810303211212, + 0.030605720356106758, + 0.02351490966975689, + 0.024986235424876213, + 0.013430009596049786, + 0.023395372554659843, + 0.021060382947325706, + 0.022208599373698235, + 0.0066590984351933, + 0.006797024980187416, + 0.0074555231258273125, + 0.004497334361076355, + 0.01000951137393713, + 0.02510184608399868, + 0.006591765210032463, + 0.01970895566046238, + 0.0038455587346106768, + 0.00407791743054986, + 0.004861140623688698, + 0.005046315491199493, + 0.006739795673638582, + 0.005118001718074083, + 0.0062343948520720005, + 0.007012601941823959, + 0.004886234644800425, + 0.006282980088144541, + 0.007241762708872557, + 0.009547973982989788, + 0.01233615167438984, + 0.005767327733337879, + 0.009341747500002384, + 0.010608496144413948, + 0.013668804429471493, + 0.01785529963672161, + 0.01502966694533825, + 0.021726341918110847, + 0.015514872036874294, + 0.016790663823485374, + 0.022152861580252647, + 0.013456753455102444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0019369428046047688 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004089090041816235 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004089090041816235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004089090041816235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12336504459381104, + 0.04666294530034065, + 0.06296773999929428, + 0.047977082431316376, + 0.057759277522563934, + 0.056876108050346375, + 0.07478214800357819, + 0.10249335318803787, + 0.27307966351509094, + 0.15403662621974945 + ], + "credit_map": [ + 0.010615833103656769, + 0.010542361065745354, + 0.028826167806982994, + 0.009962364099919796, + 0.008455546572804451, + 0.029095813632011414, + 0.013725263997912407, + 0.012141707353293896, + 0.004166174214333296, + 0.0031348303891718388, + 0.004769689869135618, + 0.0068106018006801605, + 0.005297120660543442, + 0.006255711428821087, + 0.006205886602401733, + 0.010022932663559914, + 0.006929951719939709, + 0.006769497878849506, + 0.005225407425314188, + 0.0065702698193490505, + 0.008677996695041656, + 0.004653319250792265, + 0.01606040634214878, + 0.008080894127488136, + 0.003287977073341608, + 0.0046015577390789986, + 0.005358573514968157, + 0.007386184763163328, + 0.006714791525155306, + 0.003067114157602191, + 0.011386305093765259, + 0.0061745839193463326, + 0.0071480716578662395, + 0.008352797478437424, + 0.007324459031224251, + 0.0045682876370847225, + 0.008577670902013779, + 0.007861188612878323, + 0.005742598325014114, + 0.008184211328625679, + 0.009769273921847343, + 0.009520774707198143, + 0.005827136803418398, + 0.006391020491719246, + 0.005700122565031052, + 0.006724415812641382, + 0.006059623789042234, + 0.006883746478706598, + 0.006328857038170099, + 0.0048367856070399284, + 0.005573202855885029, + 0.003460925305262208, + 0.008444639854133129, + 0.022887790575623512, + 0.007107740733772516, + 0.016142209991812706, + 0.012010090053081512, + 0.008993054740130901, + 0.011934162117540836, + 0.010385767556726933, + 0.013847656548023224, + 0.017351655289530754, + 0.014090200886130333, + 0.013880782760679722, + 0.04006393253803253, + 0.03368907794356346, + 0.03382188826799393, + 0.03503984212875366, + 0.038359642028808594, + 0.027554098516702652, + 0.037730950862169266, + 0.026820240542292595, + 0.01747642643749714, + 0.014893089421093464, + 0.01948491297662258, + 0.023671582341194153, + 0.01885213330388069, + 0.016776105388998985, + 0.025961780920624733, + 0.016920600086450577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.004089090041816235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a giant wild boar, as massive as a calf, charging forward with unstoppable force deep within an ancient, dense forest. Towering trees intertwine their branches, nearly obscuring the sky. Only a few faint, grayish-white rays of light penetrate through the gaps in the canopy, casting shifting, dappled shadows on the ground. The boar's mane, coarse, dark brown and black, is matted with damp earth and crushed plant debris. A pair of curved, yellowed tusks jut from its massive, lowered head, and its small, fierce eyes gleam with a wild glint. Each thump of its powerful hooves on the slippery, black earth studded with rotting leaves sends chunks of mud and years' worth of dead branches and leaves flying violently backward and to the sides, creating a chaotic, swirling cloud of debris visible in the dim light. It crushes small saplings and shrubs in its path. The main colors of the entire picture are deep brown and dark green, full of primitive, tense and chaotic dynamics, perfectly presenting the wild power of the beast rushing through the dense forest and being unstoppable.", + "seed": 6706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6706_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": -0.01292844833806157, + "vlm_logprob": -0.01292844833806157, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.01292844833806157, + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019268127158284187 + ], + "opposite_logprob_tokens": [ + -4.144268035888672 + ], + "logprob_reward": -0.019268127158284187, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019268127158284187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17089828848838806, + 0.045819081366062164, + 0.08470059931278229, + 0.035734739154577255, + 0.08436551690101624, + 0.07607012987136841, + 0.13195128738880157, + 0.20820258557796478, + 0.03959530219435692, + 0.1226624995470047 + ], + "credit_map": [ + 0.021131757646799088, + 0.015459195710718632, + 0.035390906035900116, + 0.018902551382780075, + 0.01132905762642622, + 0.008956300094723701, + 0.019297948107123375, + 0.04043058305978775, + 0.006723645143210888, + 0.00422592693939805, + 0.006684523541480303, + 0.009493999183177948, + 0.003543288679793477, + 0.00459462683647871, + 0.0059876092709600925, + 0.0045654610730707645, + 0.016207145527005196, + 0.009719877503812313, + 0.011510521173477173, + 0.009627102874219418, + 0.011935940943658352, + 0.006596134975552559, + 0.008851918391883373, + 0.01025195699185133, + 0.0035385119263082743, + 0.002976649673655629, + 0.0044629392214119434, + 0.008935468271374702, + 0.003475409233942628, + 0.0024575702846050262, + 0.005677280016243458, + 0.004210911691188812, + 0.007109793834388256, + 0.007453381549566984, + 0.011058232747018337, + 0.009867846965789795, + 0.01574314571917057, + 0.005108723882585764, + 0.016949540004134178, + 0.011074846610426903, + 0.011101284064352512, + 0.008476298302412033, + 0.01360688079148531, + 0.006703594233840704, + 0.010042178444564342, + 0.008488211780786514, + 0.00830644927918911, + 0.009345230646431446, + 0.020090002566576004, + 0.015815865248441696, + 0.015186073258519173, + 0.02219386212527752, + 0.013792192563414574, + 0.013211940415203571, + 0.01773422583937645, + 0.013927116058766842, + 0.01355450414121151, + 0.019905922934412956, + 0.015502860769629478, + 0.037004925310611725, + 0.04508208483457565, + 0.025525450706481934, + 0.03191814199090004, + 0.01970868557691574, + 0.002218417590484023, + 0.005159395281225443, + 0.003554859897121787, + 0.006568171549588442, + 0.005243624094873667, + 0.004863239359110594, + 0.006178643088787794, + 0.005808948539197445, + 0.009560717269778252, + 0.012653721496462822, + 0.018503224477171898, + 0.006765974685549736, + 0.02374524623155594, + 0.010263664647936821, + 0.007862739264965057, + 0.03330720588564873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.019268127158284187 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211989007890224 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014211989007890224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211989007890224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09061688184738159, + 0.08032785356044769, + 0.08287572860717773, + 0.13401353359222412, + 0.05786449462175369, + 0.04880966991186142, + 0.07714105397462845, + 0.10726298391819, + 0.05145493894815445, + 0.2696327865123749 + ], + "credit_map": [ + 0.01299633551388979, + 0.013804230839014053, + 0.011171318590641022, + 0.011503944173455238, + 0.00816540326923132, + 0.009501434862613678, + 0.012727010063827038, + 0.010747207328677177, + 0.018902691081166267, + 0.0073940688744187355, + 0.010685168206691742, + 0.009456186555325985, + 0.005497219506651163, + 0.007505251094698906, + 0.009850441478192806, + 0.011036835610866547, + 0.01298650074750185, + 0.00855686143040657, + 0.010313491337001324, + 0.011175823397934437, + 0.01285307202488184, + 0.005815483629703522, + 0.009383178316056728, + 0.01179132517427206, + 0.018254054710268974, + 0.011515708640217781, + 0.01824767515063286, + 0.03575795888900757, + 0.010330390185117722, + 0.006582967471331358, + 0.01842980459332466, + 0.014894996769726276, + 0.005305173806846142, + 0.004651723429560661, + 0.00767059251666069, + 0.008839935064315796, + 0.008155114017426968, + 0.005800149403512478, + 0.011456292122602463, + 0.005985519383102655, + 0.005160018801689148, + 0.0045838444493710995, + 0.005416841246187687, + 0.005964200012385845, + 0.008314981125295162, + 0.005245273467153311, + 0.008223049342632294, + 0.00590146891772747, + 0.014702826738357544, + 0.012051595374941826, + 0.008764595724642277, + 0.00884287990629673, + 0.007993923500180244, + 0.006147595122456551, + 0.00978772807866335, + 0.008849918842315674, + 0.00848727859556675, + 0.008681452833116055, + 0.014451498165726662, + 0.014489906840026379, + 0.013568690046668053, + 0.010717748664319515, + 0.013369307853281498, + 0.023497112095355988, + 0.0030696389731019735, + 0.007205542642623186, + 0.005499112419784069, + 0.010272146202623844, + 0.006397524382919073, + 0.003913892898708582, + 0.008023479953408241, + 0.007073608227074146, + 0.014537647366523743, + 0.017713192850351334, + 0.04291597008705139, + 0.02110653929412365, + 0.09362491220235825, + 0.04616276174783707, + 0.017080778256058693, + 0.016490992158651352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.014211989007890224 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584525540471077 + ], + "opposite_logprob_tokens": [ + -3.8955845832824707 + ], + "logprob_reward": -0.020584525540471077, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584525540471077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07063227146863937, + 0.0671096220612526, + 0.05633832514286041, + 0.07947232574224472, + 0.058345843106508255, + 0.0735815167427063, + 0.1007363498210907, + 0.15622010827064514, + 0.08140105754137039, + 0.25616249442100525 + ], + "credit_map": [ + 0.008381592109799385, + 0.009516466408967972, + 0.009744101203978062, + 0.010748172178864479, + 0.005969693884253502, + 0.005173529032617807, + 0.01297052949666977, + 0.008128187619149685, + 0.015620934776961803, + 0.006809987127780914, + 0.008799444884061813, + 0.009601014666259289, + 0.004198853857815266, + 0.005746083799749613, + 0.008719625882804394, + 0.007613681256771088, + 0.007999449968338013, + 0.004787204321473837, + 0.006706015206873417, + 0.00726864580065012, + 0.009943298995494843, + 0.0036862704437226057, + 0.011038744822144508, + 0.004908690229058266, + 0.013683831319212914, + 0.008850060403347015, + 0.008990014903247356, + 0.014063016511499882, + 0.008598042652010918, + 0.005034793168306351, + 0.011425224132835865, + 0.00882734078913927, + 0.004036553669720888, + 0.004228942561894655, + 0.008596699684858322, + 0.006885054986923933, + 0.010442613624036312, + 0.003658145433291793, + 0.01130775548517704, + 0.009190074168145657, + 0.009336346760392189, + 0.007770026568323374, + 0.010121412575244904, + 0.007979082874953747, + 0.01201239787042141, + 0.007892145775258541, + 0.009499762207269669, + 0.008970345370471478, + 0.010131402872502804, + 0.00918546412140131, + 0.011494902893900871, + 0.018244294449687004, + 0.016972171142697334, + 0.00841738935559988, + 0.013428300619125366, + 0.0128624327480793, + 0.009054180234670639, + 0.0110290152952075, + 0.012097490951418877, + 0.021376846358180046, + 0.04256771132349968, + 0.018432674929499626, + 0.027253948152065277, + 0.01440824568271637, + 0.0047369468957185745, + 0.009335367009043694, + 0.008468355983495712, + 0.019680768251419067, + 0.007877185940742493, + 0.005732339341193438, + 0.014510784298181534, + 0.011059309355914593, + 0.01665799506008625, + 0.023181308060884476, + 0.03550590202212334, + 0.01306526642292738, + 0.09030453115701675, + 0.04398515075445175, + 0.016779374331235886, + 0.016682982444763184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.020584525540471077 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12784580886363983, + 0.06966201961040497, + 0.05569012463092804, + 0.0671023577451706, + 0.09925118088722229, + 0.06234844774007797, + 0.056999798864126205, + 0.0880417600274086, + 0.037775009870529175, + 0.33528342843055725 + ], + "credit_map": [ + 0.013829047791659832, + 0.013500958681106567, + 0.023629292845726013, + 0.01773303560912609, + 0.01085078064352274, + 0.007749158889055252, + 0.01564466767013073, + 0.024908864870667458, + 0.01608302630484104, + 0.00705301109701395, + 0.008844771422445774, + 0.010945696383714676, + 0.0035145615693181753, + 0.005029913038015366, + 0.00996488519012928, + 0.008226157166063786, + 0.008946922607719898, + 0.004575211554765701, + 0.005961767863482237, + 0.006280809175223112, + 0.01057584211230278, + 0.0040132151916623116, + 0.01084946934133768, + 0.0044868867844343185, + 0.010585837997496128, + 0.006892485078424215, + 0.008070709183812141, + 0.011102715507149696, + 0.007453443948179483, + 0.00487909559160471, + 0.00945967249572277, + 0.008658397942781448, + 0.015357563272118568, + 0.01229761354625225, + 0.013135196641087532, + 0.015606311149895191, + 0.010115544311702251, + 0.007854106836020947, + 0.0122274374589324, + 0.012657403945922852, + 0.011143941432237625, + 0.0076463655568659306, + 0.009197869338095188, + 0.004734632559120655, + 0.006945313420146704, + 0.006264886353164911, + 0.006607717368751764, + 0.009807725436985493, + 0.009452775120735168, + 0.007289348170161247, + 0.007175762206315994, + 0.0069892737083137035, + 0.007123772520571947, + 0.005782275926321745, + 0.006898602936416864, + 0.006287986412644386, + 0.005389923229813576, + 0.007356089539825916, + 0.0076240068301558495, + 0.01305557694286108, + 0.022636132314801216, + 0.01225503534078598, + 0.013086892664432526, + 0.006638103164732456, + 0.0021809067111462355, + 0.004850112833082676, + 0.0045232223346829414, + 0.006503229960799217, + 0.004754518624395132, + 0.002337849233299494, + 0.007085606921464205, + 0.005539562087506056, + 0.018321597948670387, + 0.022044142708182335, + 0.054647091776132584, + 0.020667044445872307, + 0.09546133130788803, + 0.059412095695734024, + 0.02382950484752655, + 0.04090062156319618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.004635065793991089 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942534189671278 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942534189671278, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942534189671278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09270250052213669, + 0.13205158710479736, + 0.084392249584198, + 0.17500510811805725, + 0.07386184483766556, + 0.0768919438123703, + 0.08437828719615936, + 0.08585058152675629, + 0.034868642687797546, + 0.15999731421470642 + ], + "credit_map": [ + 0.012190163135528564, + 0.011506880633533001, + 0.013910267502069473, + 0.01591135375201702, + 0.007694305386394262, + 0.007544402964413166, + 0.012659759260714054, + 0.011285369284451008, + 0.025828279554843903, + 0.014064906165003777, + 0.01783142238855362, + 0.026067886501550674, + 0.0075761410407722, + 0.010382041335105896, + 0.01416336465626955, + 0.016137540340423584, + 0.016664937138557434, + 0.011462350375950336, + 0.011370384134352207, + 0.010163482278585434, + 0.008821222931146622, + 0.006719247438013554, + 0.0091076185926795, + 0.010083002969622612, + 0.028659960255026817, + 0.018359322100877762, + 0.020763661712408066, + 0.031454164534807205, + 0.018345413729548454, + 0.010290267877280712, + 0.026351818814873695, + 0.020780501887202263, + 0.013686561025679111, + 0.008272696286439896, + 0.011293765157461166, + 0.007754047401249409, + 0.00636082049459219, + 0.006399437785148621, + 0.008053038269281387, + 0.012041478417813778, + 0.010264397598803043, + 0.011923244222998619, + 0.014262867160141468, + 0.007307305466383696, + 0.009812205098569393, + 0.00552206439897418, + 0.008931625634431839, + 0.00886823795735836, + 0.012129179202020168, + 0.010470527224242687, + 0.01044287346303463, + 0.010502931661903858, + 0.00782707892358303, + 0.010410374030470848, + 0.010318955406546593, + 0.01227637194097042, + 0.0047841668128967285, + 0.006742128171026707, + 0.00914707686752081, + 0.015712406486272812, + 0.012806844897568226, + 0.009161239489912987, + 0.011330795474350452, + 0.016165925189852715, + 0.0027990301605314016, + 0.003924843855202198, + 0.003503780346363783, + 0.005861051380634308, + 0.004686511121690273, + 0.003039813134819269, + 0.005032102111726999, + 0.00602151220664382, + 0.007509416434913874, + 0.011125370860099792, + 0.01890743337571621, + 0.010167756117880344, + 0.04035646840929985, + 0.018019938841462135, + 0.010705637745559216, + 0.04320529103279114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.005942534189671278 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a giant wild boar, as massive as a calf, charging forward with unstoppable force deep within an ancient, dense forest. Towering trees intertwine their branches, nearly obscuring the sky. Only a few faint, grayish-white rays of light penetrate through the gaps in the canopy, casting shifting, dappled shadows on the ground. The boar's mane, coarse, dark brown and black, is matted with damp earth and crushed plant debris. A pair of curved, yellowed tusks jut from its massive, lowered head, and its small, fierce eyes gleam with a wild glint. Each thump of its powerful hooves on the slippery, black earth studded with rotting leaves sends chunks of mud and years' worth of dead branches and leaves flying violently backward and to the sides, creating a chaotic, swirling cloud of debris visible in the dim light. It crushes small saplings and shrubs in its path. The main colors of the entire picture are deep brown and dark green, full of primitive, tense and chaotic dynamics, perfectly presenting the wild power of the beast rushing through the dense forest and being unstoppable.", + "seed": 6707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6707_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": -0.01619065161794424, + "vlm_logprob": -0.01619065161794424, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.01619065161794424, + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0166479405015707 + ], + "opposite_logprob_tokens": [ + -4.391647815704346 + ], + "logprob_reward": -0.0166479405015707, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0166479405015707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11230304837226868, + 0.045567259192466736, + 0.04706335440278053, + 0.06087178364396095, + 0.07559727132320404, + 0.10164719074964523, + 0.07784762233495712, + 0.12090881913900375, + 0.058400724083185196, + 0.2997928559780121 + ], + "credit_map": [ + 0.012660209089517593, + 0.014930890873074532, + 0.014514628797769547, + 0.02819957584142685, + 0.00898841954767704, + 0.009249754250049591, + 0.014208217151463032, + 0.009551352821290493, + 0.0053208512254059315, + 0.005230633541941643, + 0.006238570902496576, + 0.005063403397798538, + 0.0045043593272566795, + 0.005042003467679024, + 0.004920553416013718, + 0.0092468848451972, + 0.004026421345770359, + 0.006410493049770594, + 0.005865318700671196, + 0.007543952204287052, + 0.006043025758117437, + 0.004621048923581839, + 0.005944985896348953, + 0.0066081080585718155, + 0.005529778078198433, + 0.0061056953854858875, + 0.009256693534553051, + 0.007202623877674341, + 0.011139899492263794, + 0.006466122809797525, + 0.009115253575146198, + 0.006055724341422319, + 0.009456650353968143, + 0.013657238334417343, + 0.005827227141708136, + 0.010321414098143578, + 0.009496678598225117, + 0.009597430936992168, + 0.008311850018799305, + 0.00892878882586956, + 0.00773914810270071, + 0.012736441567540169, + 0.013206982053816319, + 0.010585643351078033, + 0.009651188738644123, + 0.009532010182738304, + 0.01973382756114006, + 0.018461953848600388, + 0.006841876544058323, + 0.006469737272709608, + 0.011128036305308342, + 0.007622997276484966, + 0.012524222955107689, + 0.013929237611591816, + 0.007231604773551226, + 0.0120999151840806, + 0.007583371829241514, + 0.0083734430372715, + 0.01058975514024496, + 0.008216578513383865, + 0.017278049141168594, + 0.00933103822171688, + 0.028715305030345917, + 0.030821282416582108, + 0.0030581895262002945, + 0.007946130819618702, + 0.006471963133662939, + 0.004970660898834467, + 0.009931125678122044, + 0.006817067041993141, + 0.010214523412287235, + 0.008991065435111523, + 0.012485383078455925, + 0.028544660657644272, + 0.03673839569091797, + 0.025960471481084824, + 0.07746821641921997, + 0.023295680060982704, + 0.024265380576252937, + 0.07103469222784042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0166479405015707 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981579303741455 + ], + "opposite_logprob_tokens": [ + -3.529815673828125 + ], + "logprob_reward": -0.02981579303741455, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981579303741455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22712281346321106, + 0.07032205909490585, + 0.0815420150756836, + 0.06165298447012901, + 0.07946572452783585, + 0.11170320957899094, + 0.042623259127140045, + 0.09688350558280945, + 0.0645572617650032, + 0.1641271859407425 + ], + "credit_map": [ + 0.021670999005436897, + 0.03221278265118599, + 0.028201624751091003, + 0.050805095583200455, + 0.020760560408234596, + 0.02454197034239769, + 0.025912685319781303, + 0.023017093539237976, + 0.007675545290112495, + 0.00725184939801693, + 0.008359202183783054, + 0.01054417621344328, + 0.008360504172742367, + 0.007356744725257158, + 0.011816796846687794, + 0.008957242593169212, + 0.007789145223796368, + 0.0072212666273117065, + 0.015259812586009502, + 0.009294018149375916, + 0.007583217229694128, + 0.007072994019836187, + 0.012821394018828869, + 0.014500169083476067, + 0.004833717364817858, + 0.006524158641695976, + 0.011340891942381859, + 0.006082984618842602, + 0.00905381329357624, + 0.009730588644742966, + 0.006533146370202303, + 0.007553684990853071, + 0.0070196036249399185, + 0.011411884799599648, + 0.006425343453884125, + 0.01111380010843277, + 0.012543516233563423, + 0.012012233026325703, + 0.010256798937916756, + 0.008682546205818653, + 0.01031961664557457, + 0.011423875577747822, + 0.013970090076327324, + 0.012586530297994614, + 0.011827275156974792, + 0.009355880320072174, + 0.02168351039290428, + 0.020536430180072784, + 0.004430097993463278, + 0.0036804405972361565, + 0.005539645906537771, + 0.004386274144053459, + 0.0072575947269797325, + 0.007917575538158417, + 0.003541564801707864, + 0.005870064254850149, + 0.0071830302476882935, + 0.0062618558295071125, + 0.010473703034222126, + 0.006041223183274269, + 0.014897963963449001, + 0.007789378985762596, + 0.018691286444664, + 0.025545060634613037, + 0.004778818693011999, + 0.00781329721212387, + 0.007618119940161705, + 0.005634544882923365, + 0.010471788235008717, + 0.007174634840339422, + 0.009876220487058163, + 0.011189838871359825, + 0.007299827411770821, + 0.01972665637731552, + 0.024049928411841393, + 0.018134573474526405, + 0.037430379539728165, + 0.015274107456207275, + 0.015930788591504097, + 0.026280928403139114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02981579303741455 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078105196356773 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078105196356773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078105196356773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17702807486057281, + 0.0845925435423851, + 0.07505922764539719, + 0.2974957823753357, + 0.0508534200489521, + 0.05106779560446739, + 0.12071441113948822, + 0.02077253721654415, + 0.039395250380039215, + 0.0830209031701088 + ], + "credit_map": [ + 0.01949111744761467, + 0.024537350982427597, + 0.024085793644189835, + 0.04346771910786629, + 0.015437542460858822, + 0.014981902204453945, + 0.02048233151435852, + 0.014544333331286907, + 0.007649621460586786, + 0.011357736773788929, + 0.010430210269987583, + 0.012370600365102291, + 0.007586849387735128, + 0.009890202432870865, + 0.013308691792190075, + 0.01199864037334919, + 0.005708782933652401, + 0.01159807201474905, + 0.010291103273630142, + 0.01171097345650196, + 0.010546589270234108, + 0.006160295568406582, + 0.01050654798746109, + 0.008536869660019875, + 0.025578806176781654, + 0.020574679598212242, + 0.0643487349152565, + 0.02787071280181408, + 0.0365440659224987, + 0.043957460671663284, + 0.04915847256779671, + 0.029462873935699463, + 0.007939331233501434, + 0.007489298935979605, + 0.004656658507883549, + 0.006389436777681112, + 0.006357552018016577, + 0.006533448584377766, + 0.005269635934382677, + 0.00621806038543582, + 0.00351018738001585, + 0.0062273344956338406, + 0.007118791341781616, + 0.005897727329283953, + 0.0051468596793711185, + 0.004627020098268986, + 0.009624920785427094, + 0.008914961479604244, + 0.008969474583864212, + 0.007830602116882801, + 0.03106771409511566, + 0.009832082316279411, + 0.024987366050481796, + 0.015630099922418594, + 0.009925255551934242, + 0.012471823953092098, + 0.0015847717877477407, + 0.0025801509618759155, + 0.002510861959308386, + 0.0016065145609900355, + 0.0036135397385805845, + 0.0022355136461555958, + 0.0033190425019711256, + 0.0033221442718058825, + 0.0026992489583790302, + 0.004764165263622999, + 0.004280403722077608, + 0.0037226080894470215, + 0.00658416235819459, + 0.0045197708532214165, + 0.006256284657865763, + 0.006568609271198511, + 0.005527602043002844, + 0.010073070414364338, + 0.010237847454845905, + 0.009630000218749046, + 0.015084223821759224, + 0.010075950063765049, + 0.008867735974490643, + 0.013524478301405907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.016078105196356773 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636386133730412, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636386133730412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1564897894859314, + 0.15275748074054718, + 0.03824342042207718, + 0.06043408066034317, + 0.019682198762893677, + 0.07018331438302994, + 0.10007365792989731, + 0.09155292063951492, + 0.06304412335157394, + 0.24753901362419128 + ], + "credit_map": [ + 0.018555693328380585, + 0.023522470146417618, + 0.02138819918036461, + 0.03095938451588154, + 0.013841694220900536, + 0.017351988703012466, + 0.016726210713386536, + 0.014144151471555233, + 0.01669834926724434, + 0.023358413949608803, + 0.01999431662261486, + 0.019293460994958878, + 0.014848110266029835, + 0.020179998129606247, + 0.0185999795794487, + 0.019784845411777496, + 0.005850962828844786, + 0.0032289260998368263, + 0.004462643060833216, + 0.002647882327437401, + 0.004614970181137323, + 0.008184039033949375, + 0.003608588594943285, + 0.00564540782943368, + 0.007015864364802837, + 0.0053276545368134975, + 0.009536783210933208, + 0.004862470552325249, + 0.00869495514780283, + 0.008357163518667221, + 0.006783905904740095, + 0.009855283424258232, + 0.0020835353061556816, + 0.002575768856331706, + 0.0017792056314647198, + 0.002105985302478075, + 0.0027860451955348253, + 0.0031386022455990314, + 0.00212970282882452, + 0.0030833526980131865, + 0.005941556300967932, + 0.008296030573546886, + 0.00957390945404768, + 0.006910740397870541, + 0.007524935994297266, + 0.008914503268897533, + 0.011608210392296314, + 0.01141342893242836, + 0.005565920379012823, + 0.007608401123434305, + 0.019345518201589584, + 0.007851087488234043, + 0.016513891518115997, + 0.017296701669692993, + 0.00867510586977005, + 0.017217038199305534, + 0.005411833990365267, + 0.008483164943754673, + 0.007940090261399746, + 0.005519603379070759, + 0.014519715681672096, + 0.008412466384470463, + 0.017155425623059273, + 0.024110618978738785, + 0.002723476616665721, + 0.006551878992468119, + 0.0074223787523806095, + 0.008490733802318573, + 0.011108595877885818, + 0.00510805007070303, + 0.012307551689445972, + 0.00933146197348833, + 0.010393750853836536, + 0.026666944846510887, + 0.027860024943947792, + 0.019454345107078552, + 0.0639660656452179, + 0.01934552937746048, + 0.017284220084547997, + 0.06256812065839767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008636386133730412 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775033220648766 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775033220648766, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775033220648766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11916258931159973, + 0.10436028242111206, + 0.04576783999800682, + 0.1882188469171524, + 0.0972636416554451, + 0.042547523975372314, + 0.05907580628991127, + 0.10718301683664322, + 0.06716971844434738, + 0.16925077140331268 + ], + "credit_map": [ + 0.01761036552488804, + 0.012461090460419655, + 0.017539067193865776, + 0.01011891569942236, + 0.012033050879836082, + 0.014333766885101795, + 0.014536097645759583, + 0.020530223846435547, + 0.010366672649979591, + 0.013239564374089241, + 0.014541867189109325, + 0.015932751819491386, + 0.008979056030511856, + 0.013494361191987991, + 0.014223401434719563, + 0.01358259841799736, + 0.005775881465524435, + 0.004318500403314829, + 0.004727763123810291, + 0.004185079596936703, + 0.00664690975099802, + 0.006870705168694258, + 0.006010915152728558, + 0.007232083007693291, + 0.019221005961298943, + 0.012935095466673374, + 0.05172811821103096, + 0.009625755250453949, + 0.021540600806474686, + 0.03518964350223541, + 0.02444429323077202, + 0.013534320518374443, + 0.014829035848379135, + 0.013385231606662273, + 0.009436385706067085, + 0.012899194844067097, + 0.009775146842002869, + 0.0133128110319376, + 0.011086579412221909, + 0.012539250776171684, + 0.005507313180714846, + 0.004073369782418013, + 0.006833911407738924, + 0.004637653939425945, + 0.0036398593802005053, + 0.006700150202959776, + 0.004826126620173454, + 0.0063291373662650585, + 0.005022940691560507, + 0.005208821967244148, + 0.009922360070049763, + 0.005260947626084089, + 0.00950678437948227, + 0.009657785296440125, + 0.005660676397383213, + 0.008835487067699432, + 0.014046557247638702, + 0.013282651081681252, + 0.01729431003332138, + 0.0068251569755375385, + 0.015817683190107346, + 0.012574025429785252, + 0.008446188643574715, + 0.018896428868174553, + 0.004146726336330175, + 0.007119089365005493, + 0.0074593001045286655, + 0.00821429118514061, + 0.011403732933104038, + 0.005320690106600523, + 0.01178512629121542, + 0.011720758862793446, + 0.010097539983689785, + 0.016546746715903282, + 0.02302497625350952, + 0.013911472633481026, + 0.029933493584394455, + 0.015962298959493637, + 0.010107372887432575, + 0.0496668703854084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.009775033220648766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "subsequent low-angle, dynamic tracking shot captures a giant wild boar, as massive as a calf, charging forward with unstoppable force deep within an ancient, dense forest. Towering trees intertwine their branches, nearly obscuring the sky. Only a few faint, grayish-white rays of light penetrate through the gaps in the canopy, casting shifting, dappled shadows on the ground. The boar's mane, coarse, dark brown and black, is matted with damp earth and crushed plant debris. A pair of curved, yellowed tusks jut from its massive, lowered head, and its small, fierce eyes gleam with a wild glint. Each thump of its powerful hooves on the slippery, black earth studded with rotting leaves sends chunks of mud and years' worth of dead branches and leaves flying violently backward and to the sides, creating a chaotic, swirling cloud of debris visible in the dim light. It crushes small saplings and shrubs in its path. The main colors of the entire picture are deep brown and dark green, full of primitive, tense and chaotic dynamics, perfectly presenting the wild power of the beast rushing through the dense forest and being unstoppable.", + "seed": 6708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_6/index_3324_seed_6708_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": -0.01902284426614642, + "vlm_logprob": -0.01902284426614642, + "vlm_yes_no_margin": 4.224999952316284, + "vlm_token_credit": -0.01902284426614642, + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037306107580661774 + ], + "opposite_logprob_tokens": [ + -3.412306070327759 + ], + "logprob_reward": -0.037306107580661774, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037306107580661774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19694235920906067, + 0.030566206201910973, + 0.12588395178318024, + 0.062000442296266556, + 0.04999719560146332, + 0.0499175488948822, + 0.23221226036548615, + 0.10851164162158966, + 0.0575392059981823, + 0.08642913401126862 + ], + "credit_map": [ + 0.020465515553951263, + 0.029338644817471504, + 0.028115754947066307, + 0.022894613444805145, + 0.02860194258391857, + 0.01755347102880478, + 0.0330679677426815, + 0.016904491931200027, + 0.003396948566660285, + 0.0030728920828551054, + 0.004595567937940359, + 0.0064675649628043175, + 0.0030067383777350187, + 0.0025463455822318792, + 0.004743367899209261, + 0.002736784750595689, + 0.011609937064349651, + 0.04118812829256058, + 0.014494691044092178, + 0.013580250553786755, + 0.010333902202546597, + 0.008409104309976101, + 0.01317867636680603, + 0.01308927871286869, + 0.006083320360630751, + 0.010949155315756798, + 0.008141644299030304, + 0.007864422164857388, + 0.010855010710656643, + 0.004184162709861994, + 0.00641421414911747, + 0.007508526090532541, + 0.004680424463003874, + 0.004759168718010187, + 0.007433761842548847, + 0.004534335341304541, + 0.008744297549128532, + 0.0057213096879422665, + 0.008470184169709682, + 0.005653719883412123, + 0.003906530793756247, + 0.00693926028907299, + 0.006715062074363232, + 0.006949060596525669, + 0.010050077922642231, + 0.0035125284921377897, + 0.007266814820468426, + 0.004578218795359135, + 0.017627455294132233, + 0.020203273743391037, + 0.027651332318782806, + 0.04045947268605232, + 0.027130059897899628, + 0.017202945426106453, + 0.05278987064957619, + 0.029147867113351822, + 0.011473482474684715, + 0.009053302928805351, + 0.01410052552819252, + 0.02216687612235546, + 0.010151981376111507, + 0.013738816604018211, + 0.01706908829510212, + 0.01075758133083582, + 0.0032942777033895254, + 0.005703788250684738, + 0.006429817993193865, + 0.009071292355656624, + 0.014101735316216946, + 0.003836426418274641, + 0.0073777781799435616, + 0.0077240951359272, + 0.006817047484219074, + 0.008215452544391155, + 0.011560159735381603, + 0.015002180822193623, + 0.010815337300300598, + 0.006843331269919872, + 0.014665885828435421, + 0.012509742751717567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.037306107580661774 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635374292731285 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.02635374292731285, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635374292731285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0849723294377327, + 0.043339796364307404, + 0.08903292566537857, + 0.08305640518665314, + 0.026932524517178535, + 0.06014744192361832, + 0.1574552059173584, + 0.23803824186325073, + 0.09062033891677856, + 0.12640486657619476 + ], + "credit_map": [ + 0.008337428793311119, + 0.011008876375854015, + 0.011336038820445538, + 0.016677848994731903, + 0.008350684307515621, + 0.009244517423212528, + 0.010070128366351128, + 0.009946804493665695, + 0.0047610606998205185, + 0.004423600155860186, + 0.006490906700491905, + 0.011251832358539104, + 0.0036012062337249517, + 0.0029942449182271957, + 0.005745348986238241, + 0.004071593750268221, + 0.008647508919239044, + 0.009893382899463177, + 0.010725964792072773, + 0.009132466278970242, + 0.019251449033617973, + 0.008140776306390762, + 0.015003375709056854, + 0.008237987756729126, + 0.008694526739418507, + 0.014621459878981113, + 0.010488555766642094, + 0.012220474891364574, + 0.010789770632982254, + 0.005392162594944239, + 0.01125976163893938, + 0.009589691646397114, + 0.0027891648933291435, + 0.001668061246164143, + 0.004398365505039692, + 0.0033543496392667294, + 0.0034087705425918102, + 0.0018750113667920232, + 0.006161733064800501, + 0.003277064301073551, + 0.006611214019358158, + 0.006742982193827629, + 0.008216590620577335, + 0.010896945372223854, + 0.007316181901842356, + 0.0061827716417610645, + 0.00783579982817173, + 0.006344947963953018, + 0.014018838293850422, + 0.01421886682510376, + 0.022551745176315308, + 0.03161254897713661, + 0.01813993602991104, + 0.01292975340038538, + 0.02715679444372654, + 0.016826709732413292, + 0.02214881405234337, + 0.018982723355293274, + 0.024414053186774254, + 0.04208412021398544, + 0.026397349312901497, + 0.021012861281633377, + 0.05840037763118744, + 0.024597927927970886, + 0.007875987328588963, + 0.008045762777328491, + 0.011546882800757885, + 0.010791764594614506, + 0.01781255193054676, + 0.006262301933020353, + 0.015683885663747787, + 0.012601188383996487, + 0.009111076593399048, + 0.008951568976044655, + 0.016037188470363617, + 0.020580409094691277, + 0.012365616858005524, + 0.01707252860069275, + 0.02232104167342186, + 0.019965434446930885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.02635374292731285 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188709393143654 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188709393143654, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188709393143654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03439173847436905, + 0.019996637478470802, + 0.02089294232428074, + 0.03727250173687935, + 0.01907523162662983, + 0.016560185700654984, + 0.10951781272888184, + 0.6663243174552917, + 0.039359550923109055, + 0.03660902753472328 + ], + "credit_map": [ + 0.004135231953114271, + 0.004558234475553036, + 0.005809072870761156, + 0.0036032095085829496, + 0.004725890699774027, + 0.00360794342122972, + 0.005474413279443979, + 0.0024777462240308523, + 0.0021668625995516777, + 0.0017648927168920636, + 0.003156829159706831, + 0.005059588234871626, + 0.002036296995356679, + 0.0011378638446331024, + 0.0029110778123140335, + 0.0017632270464673638, + 0.0023421174846589565, + 0.0025946765672415495, + 0.002818843349814415, + 0.0025949596893042326, + 0.0032367596868425608, + 0.0014611531514674425, + 0.003939576912671328, + 0.0019048562971875072, + 0.003828555578365922, + 0.005095072090625763, + 0.005977672524750233, + 0.004597177729010582, + 0.006305365357547998, + 0.0027521043084561825, + 0.004768102429807186, + 0.003948458004742861, + 0.0019121433142572641, + 0.002113988157361746, + 0.0030396138317883015, + 0.0016982117667794228, + 0.0028474698774516582, + 0.001975318416953087, + 0.003342845244333148, + 0.0021456428803503513, + 0.001521365367807448, + 0.002339667174965143, + 0.0022580940276384354, + 0.002331350464373827, + 0.0024366346187889576, + 0.0012753038899973035, + 0.0025210503954440355, + 0.0018767203437164426, + 0.008594531565904617, + 0.008993711322546005, + 0.011719657108187675, + 0.021318115293979645, + 0.012878911569714546, + 0.00886419415473938, + 0.02276156097650528, + 0.014387144707143307, + 0.07032160460948944, + 0.06487178057432175, + 0.07267789542675018, + 0.1088443472981453, + 0.06993836909532547, + 0.06178313493728638, + 0.1509486585855484, + 0.06693858653306961, + 0.0019742080476135015, + 0.004654718562960625, + 0.0044175186194479465, + 0.005300475750118494, + 0.007453613914549351, + 0.003741438966244459, + 0.007252085953950882, + 0.004565494135022163, + 0.0024179844185709953, + 0.0028403019532561302, + 0.005655474495142698, + 0.008274638094007969, + 0.003765104105696082, + 0.0034233739133924246, + 0.005474795587360859, + 0.004757355432957411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.018188709393143654 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635184537619352 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635184537619352, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635184537619352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13420166075229645, + 0.06909547746181488, + 0.1424660086631775, + 0.02579600363969803, + 0.09004687517881393, + 0.020513948053121567, + 0.04286159202456474, + 0.3646494150161743, + 0.06376785784959793, + 0.046601131558418274 + ], + "credit_map": [ + 0.017282580956816673, + 0.018404463306069374, + 0.025337470695376396, + 0.018390022218227386, + 0.013642757199704647, + 0.013465967029333115, + 0.014057072810828686, + 0.01362135075032711, + 0.013122962787747383, + 0.006070608738809824, + 0.016416048631072044, + 0.009029124863445759, + 0.007931713946163654, + 0.005294926930218935, + 0.00571323512122035, + 0.0055168624967336655, + 0.014528140425682068, + 0.05065303295850754, + 0.015398530289530754, + 0.012078426778316498, + 0.013424338772892952, + 0.009105289354920387, + 0.013388658873736858, + 0.013889596797525883, + 0.0031401473097503185, + 0.0038068124558776617, + 0.003973650746047497, + 0.0031019835732877254, + 0.002416388364508748, + 0.0021302986424416304, + 0.00355306314304471, + 0.0036736619658768177, + 0.009199092164635658, + 0.010204005055129528, + 0.015172130428254604, + 0.015419634990394115, + 0.011913097463548183, + 0.0069160498678684235, + 0.011724241077899933, + 0.00949863437563181, + 0.0018615934532135725, + 0.0028728318866342306, + 0.0029020782094448805, + 0.002487510209903121, + 0.0039345272816717625, + 0.001410724245943129, + 0.0026219862047582865, + 0.002422698773443699, + 0.003808744717389345, + 0.004697612952440977, + 0.005756918806582689, + 0.0071838414296507835, + 0.006582781672477722, + 0.003999859094619751, + 0.006654351018369198, + 0.004177485592663288, + 0.036640629172325134, + 0.03604679927229881, + 0.039328377693891525, + 0.060243237763643265, + 0.03845813497900963, + 0.031659264117479324, + 0.08374819904565811, + 0.03852478414773941, + 0.0027146798092871904, + 0.004331903997808695, + 0.004467765800654888, + 0.0068155815824866295, + 0.018791137263178825, + 0.0039872657507658005, + 0.00663352245464921, + 0.016026010736823082, + 0.0020815348252654076, + 0.003559742122888565, + 0.004456060007214546, + 0.006452231667935848, + 0.00726326135918498, + 0.008574901148676872, + 0.006528515834361315, + 0.007684891577810049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004635184537619352 + }, + { + "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 giant wild boar with a mane of coarse dark brown and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar charging forward with powerful hooves that kick up mud, dead branches, and leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the boar's curved yellowed tusks and small fierce eyes visible on its lowered head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment an ancient dense forest with towering trees and faint grayish-white rays of light penetrating the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle dynamic tracking shot following the boar's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630476891994476 + ], + "opposite_logprob_tokens": [ + -4.758630275726318 + ], + "logprob_reward": -0.008630476891994476, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630476891994476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07916814833879471, + 0.024965792894363403, + 0.0845869779586792, + 0.0365934744477272, + 0.05142100155353546, + 0.05618011951446533, + 0.08256097137928009, + 0.32791122794151306, + 0.08161556720733643, + 0.17499668896198273 + ], + "credit_map": [ + 0.008212562650442123, + 0.011528022587299347, + 0.010373895056545734, + 0.016483398154377937, + 0.00860634632408619, + 0.0062539177015423775, + 0.010485431179404259, + 0.007224574685096741, + 0.0023965167347341776, + 0.0025923210196197033, + 0.003728394163772464, + 0.006794071290642023, + 0.00255829025991261, + 0.001940761343576014, + 0.0027654569130390882, + 0.0021899831481277943, + 0.008477152325212955, + 0.023011906072497368, + 0.009303286671638489, + 0.008113649673759937, + 0.010854251682758331, + 0.005918728653341532, + 0.010032215155661106, + 0.008875787258148193, + 0.004727873485535383, + 0.005393016617745161, + 0.005411562044173479, + 0.004056869074702263, + 0.00554023589938879, + 0.002591403666883707, + 0.004529409110546112, + 0.004343101754784584, + 0.005197536200284958, + 0.003927102778106928, + 0.007844362407922745, + 0.006092843133956194, + 0.010283630341291428, + 0.004118069540709257, + 0.00950394943356514, + 0.004453509580343962, + 0.004471250809729099, + 0.005931244231760502, + 0.008530085906386375, + 0.008462523110210896, + 0.010773437097668648, + 0.00442951126024127, + 0.008650318719446659, + 0.004931748379021883, + 0.006923047825694084, + 0.005837078206241131, + 0.01217015739530325, + 0.01723666861653328, + 0.008112234063446522, + 0.006258627865463495, + 0.015088568441569805, + 0.01093458104878664, + 0.030302181839942932, + 0.03573693335056305, + 0.045325230807065964, + 0.05553434044122696, + 0.03758775442838669, + 0.026352718472480774, + 0.06632746756076813, + 0.030744601041078568, + 0.005293083842843771, + 0.00515751400962472, + 0.007312898989766836, + 0.01095998752862215, + 0.023157397285103798, + 0.004406728781759739, + 0.008787773549556732, + 0.016540180891752243, + 0.016164755448698997, + 0.017041359096765518, + 0.025922369211912155, + 0.032970331609249115, + 0.015048500150442123, + 0.011716371402144432, + 0.030186591669917107, + 0.02594640664756298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.008630476891994476 + } + ], + "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/0000060/group_7/index_tensor([6082, 6082])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_tensor([6082, 6082])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8cca88280c63da836587b7543950a304e60a7f42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_tensor([6082, 6082])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, high-speed aerial shot, mimicking the style of an Olympic live broadcast, tilts downward and pans sideways to capture a heart-pounding chase. In the frame, a legendary tengu swoops down from the summit of a jagged, snow-capped mountain, his form dwarfed yet explosive against the vast, deep blue sky. With a crimson face and a signature, unusually long nose, his expression is focused and fierce, his gaze fixed like a hawk on a distant target in the valley below. He clads himself in a pair of massive, jet-black wings, each powerful flap seeming to tear through the clear, crisp mountain air. His body, tensed into a perfectly streamlined form, impacts the ground with such incredible force and speed that even the air ripples subtly around him. The traditional Yamabushi ascetic clothing he wore danced wildly in the high-speed airflow. The whole scene, illuminated by the bright and dazzling sunlight, between the vast blue sky and the continuous snow-capped peaks, created a grand, primitive and extremely competitive atmosphere, perfectly demonstrating the absolute power and speed of this mythical creature.", + "seed": 11620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11620_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": -0.020839165640063585, + "vlm_logprob": -0.020839165640063585, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.020839165640063585, + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034329865127801895 + ], + "opposite_logprob_tokens": [ + -3.409329891204834 + ], + "logprob_reward": -0.034329865127801895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.034329865127801895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3068096935749054, + 0.07767163962125778, + 0.06141141802072525, + 0.020158441737294197, + 0.04450435936450958, + 0.02689582109451294, + 0.1762048453092575, + 0.0770164430141449, + 0.04627499356865883, + 0.16305239498615265 + ], + "credit_map": [ + 0.03798946365714073, + 0.07395825535058975, + 0.03657824546098709, + 0.04704393818974495, + 0.03568950667977333, + 0.02456102892756462, + 0.026371009647846222, + 0.024618225172162056, + 0.00831567496061325, + 0.009199380874633789, + 0.012195231392979622, + 0.0092292670160532, + 0.0077769141644239426, + 0.007855077274143696, + 0.011551233939826488, + 0.011548848822712898, + 0.009696568362414837, + 0.006371795199811459, + 0.005587239284068346, + 0.008270727470517159, + 0.007178695406764746, + 0.010519083589315414, + 0.008974720723927021, + 0.00481257913634181, + 0.0026537447702139616, + 0.0029631757643073797, + 0.0020764111541211605, + 0.002315589925274253, + 0.0021251519210636616, + 0.002634711330756545, + 0.002372737042605877, + 0.003016915637999773, + 0.005768719129264355, + 0.009102419950067997, + 0.003053106600418687, + 0.005982712842524052, + 0.006549241952598095, + 0.00440954277291894, + 0.005832853727042675, + 0.003805754939094186, + 0.0034194502513855696, + 0.00255176005885005, + 0.004374073352664709, + 0.002969834255054593, + 0.002682524034753442, + 0.0034857969731092453, + 0.004010826349258423, + 0.003401552326977253, + 0.02021954208612442, + 0.015823902562260628, + 0.02245701290667057, + 0.015200740657746792, + 0.032513685524463654, + 0.01326264999806881, + 0.03504053130745888, + 0.021686764433979988, + 0.011394062079489231, + 0.008161192759871483, + 0.012309596873819828, + 0.008811176754534245, + 0.008407710120081902, + 0.006167819257825613, + 0.011546393856406212, + 0.010218481533229351, + 0.005821108352392912, + 0.004793819971382618, + 0.004604435991495848, + 0.005061259493231773, + 0.0072407531552016735, + 0.0045815384946763515, + 0.008292583748698235, + 0.005879489704966545, + 0.022833602502942085, + 0.03384306654334068, + 0.020063959062099457, + 0.015723131597042084, + 0.023138215765357018, + 0.013681048527359962, + 0.02225743606686592, + 0.011511918157339096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.034329865127801895 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.28282395005226135, + 0.1089939996600151, + 0.10672419518232346, + 0.05257457122206688, + 0.04315360635519028, + 0.028944477438926697, + 0.07635179907083511, + 0.05243315175175667, + 0.07876625657081604, + 0.16923397779464722 + ], + "credit_map": [ + 0.032947346568107605, + 0.05934209004044533, + 0.031114324927330017, + 0.061611391603946686, + 0.03396005555987358, + 0.021756162866950035, + 0.02250341698527336, + 0.01958915777504444, + 0.012380767613649368, + 0.013799754902720451, + 0.021908599883317947, + 0.015561290085315704, + 0.01085310522466898, + 0.0110813919454813, + 0.011867559514939785, + 0.011541529558598995, + 0.012759674340486526, + 0.009608953259885311, + 0.01096819993108511, + 0.019555089995265007, + 0.011411932297050953, + 0.016482915729284286, + 0.013476334512233734, + 0.01246109139174223, + 0.006788794416934252, + 0.0076999422162771225, + 0.007848162204027176, + 0.006066600326448679, + 0.007410428952425718, + 0.006602781359106302, + 0.005892787594348192, + 0.004265073221176863, + 0.006780983414500952, + 0.010963612236082554, + 0.0032590271439403296, + 0.005249780602753162, + 0.00568536389619112, + 0.003953520208597183, + 0.0037740461993962526, + 0.0034872700925916433, + 0.00349336676299572, + 0.002283295150846243, + 0.0042577143758535385, + 0.003504628548398614, + 0.003195335390046239, + 0.004034122917801142, + 0.004814059939235449, + 0.0033619541209191084, + 0.008283942006528378, + 0.007902806624770164, + 0.009603499434888363, + 0.006955923046916723, + 0.01219622790813446, + 0.006429848726838827, + 0.01433493196964264, + 0.010644626803696156, + 0.007230149116367102, + 0.004638828802853823, + 0.007181211374700069, + 0.005597974173724651, + 0.00529094971716404, + 0.004437265917658806, + 0.011687561869621277, + 0.006369207985699177, + 0.011897321790456772, + 0.01397307962179184, + 0.007856284268200397, + 0.014032394625246525, + 0.008021386340260506, + 0.008776620030403137, + 0.008278791792690754, + 0.00593037623912096, + 0.025068145245313644, + 0.037978678941726685, + 0.0222768597304821, + 0.01681334711611271, + 0.02127907983958721, + 0.013372544199228287, + 0.021816637367010117, + 0.010628685355186462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008647376671433449 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178047270514071 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.001178047270514071, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178047270514071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.430353045463562, + 0.10070307552814484, + 0.06493842601776123, + 0.050934724509716034, + 0.03692038357257843, + 0.03082231618463993, + 0.049570366740226746, + 0.013666262850165367, + 0.036121729761362076, + 0.18596962094306946 + ], + "credit_map": [ + 0.0606878437101841, + 0.16992560029029846, + 0.01311703771352768, + 0.022147836163640022, + 0.11395944654941559, + 0.015306299552321434, + 0.01663285680115223, + 0.018576156347990036, + 0.013636005111038685, + 0.011852454394102097, + 0.014675588347017765, + 0.012765898369252682, + 0.01094533409923315, + 0.010686878114938736, + 0.015457878820598125, + 0.010683044791221619, + 0.008100026287138462, + 0.006780771538615227, + 0.0033313382882624865, + 0.005694705061614513, + 0.0077657923102378845, + 0.010580988600850105, + 0.01377013698220253, + 0.00891466997563839, + 0.006202855613082647, + 0.007914901711046696, + 0.004916569218039513, + 0.004182189702987671, + 0.006605919916182756, + 0.00910621602088213, + 0.006676053162664175, + 0.005330022424459457, + 0.00489226495847106, + 0.006230928935110569, + 0.0031992727890610695, + 0.0047592176124453545, + 0.0032190782949328423, + 0.005471276585012674, + 0.00460068741813302, + 0.004547660239040852, + 0.004125769715756178, + 0.0029233393725007772, + 0.0034728683531284332, + 0.003698880085721612, + 0.004103904124349356, + 0.0033664570655673742, + 0.005489477887749672, + 0.0036416228394955397, + 0.009470799006521702, + 0.009571796283125877, + 0.004278849810361862, + 0.0038792849518358707, + 0.0042990995571017265, + 0.005156734026968479, + 0.005935043096542358, + 0.006978761870414019, + 0.0026530034374445677, + 0.0016706521855667233, + 0.0027584454510360956, + 0.001199244987219572, + 0.0012893177336081862, + 0.001284578233025968, + 0.001387575059197843, + 0.001423446461558342, + 0.0061009954661130905, + 0.006456385832279921, + 0.0037896824069321156, + 0.006459589581936598, + 0.003037342568859458, + 0.0046059065498411655, + 0.0032099506352096796, + 0.002461880212649703, + 0.02807152457535267, + 0.04161393642425537, + 0.02632906474173069, + 0.020287267863750458, + 0.022284822538495064, + 0.014532902278006077, + 0.022425368428230286, + 0.01042473316192627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.001178047270514071 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263367909938097 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.0263367909938097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263367909938097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20960953831672668, + 0.10459436476230621, + 0.051781803369522095, + 0.02709135226905346, + 0.04440377652645111, + 0.05065365880727768, + 0.12549497187137604, + 0.038317471742630005, + 0.07493294775485992, + 0.2731201648712158 + ], + "credit_map": [ + 0.02506340853869915, + 0.05980930104851723, + 0.015422474592924118, + 0.02151874452829361, + 0.04167880862951279, + 0.01409912109375, + 0.014420567080378532, + 0.01759711466729641, + 0.012553734704852104, + 0.012619107030332088, + 0.019667500630021095, + 0.0134093277156353, + 0.010032309219241142, + 0.010696295648813248, + 0.016150306910276413, + 0.009465772658586502, + 0.006941735744476318, + 0.0062850164249539375, + 0.004742792341858149, + 0.008047145791351795, + 0.004774273373186588, + 0.0075621576979756355, + 0.0077863894402980804, + 0.0056422860361635685, + 0.0037449582014232874, + 0.004370293579995632, + 0.0037763421423733234, + 0.0031182379461824894, + 0.0038190986961126328, + 0.003635973669588566, + 0.0024958199355751276, + 0.0021306260023266077, + 0.007878664880990982, + 0.007670514285564423, + 0.005540973506867886, + 0.004937157034873962, + 0.0060368869453668594, + 0.00451022619381547, + 0.004205855540931225, + 0.0036234946455806494, + 0.0071136243641376495, + 0.004732009023427963, + 0.005163331050425768, + 0.0070456573739647865, + 0.006156090646982193, + 0.004865008406341076, + 0.0089516406878829, + 0.006626292131841183, + 0.019027864560484886, + 0.01542019471526146, + 0.01111568696796894, + 0.009818041697144508, + 0.01150684803724289, + 0.01755661703646183, + 0.015174676664173603, + 0.02587503381073475, + 0.007172053679823875, + 0.00477714603766799, + 0.007510669995099306, + 0.0036353867035359144, + 0.0030791719909757376, + 0.003838383359834552, + 0.004161986988037825, + 0.004142668563872576, + 0.013670484535396099, + 0.013419998809695244, + 0.007866574451327324, + 0.009266166016459465, + 0.010579163208603859, + 0.008291477337479591, + 0.006972291506826878, + 0.004866787698119879, + 0.04224243760108948, + 0.06585557013750076, + 0.03316633030772209, + 0.028129011392593384, + 0.03507018834352493, + 0.020615646615624428, + 0.032625164836645126, + 0.0154157979413867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0263367909938097 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033703748136758804 + ], + "opposite_logprob_tokens": [ + -3.4087038040161133 + ], + "logprob_reward": -0.033703748136758804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033703748136758804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5261164903640747, + 0.05019338056445122, + 0.053762972354888916, + 0.04412398487329483, + 0.036703843623399734, + 0.03436293452978134, + 0.0982026681303978, + 0.03456811234354973, + 0.07931763678789139, + 0.04264790937304497 + ], + "credit_map": [ + 0.07186485081911087, + 0.20166894793510437, + 0.021262243390083313, + 0.025532180443406105, + 0.13716498017311096, + 0.02212102897465229, + 0.02415875345468521, + 0.022343527525663376, + 0.004697831813246012, + 0.005158842541277409, + 0.004102551843971014, + 0.004951287992298603, + 0.005511700641363859, + 0.0061056953854858875, + 0.009488691575825214, + 0.01017677877098322, + 0.006519986782222986, + 0.005943437572568655, + 0.003913547843694687, + 0.004983184859156609, + 0.0044769286178052425, + 0.010198581963777542, + 0.010461107827723026, + 0.007266197819262743, + 0.00646762503311038, + 0.006100373342633247, + 0.003915087319910526, + 0.005953049287199974, + 0.00537265045568347, + 0.005287175066769123, + 0.005566202569752932, + 0.005461822729557753, + 0.00482666352763772, + 0.006407266948372126, + 0.0032317801378667355, + 0.005787495989352465, + 0.006335435435175896, + 0.00454701529815793, + 0.0029335464350879192, + 0.0026346405502408743, + 0.005107705015689135, + 0.0028596040792763233, + 0.00471569923684001, + 0.0036881761625409126, + 0.004271961748600006, + 0.0036869917530566454, + 0.005942476447671652, + 0.004090319387614727, + 0.014306887984275818, + 0.014013184234499931, + 0.011640035547316074, + 0.009735641069710255, + 0.010306404903531075, + 0.012431851588189602, + 0.012841315008699894, + 0.012927351519465446, + 0.007400038186460733, + 0.005489703733474016, + 0.006110512185841799, + 0.0035823022481054068, + 0.0025176657363772392, + 0.003258008975535631, + 0.0031299139373004436, + 0.003079967573285103, + 0.015078538097441196, + 0.01622239127755165, + 0.007222544867545366, + 0.01154374796897173, + 0.007930116727948189, + 0.010032243095338345, + 0.006021585781127214, + 0.00526647362858057, + 0.006628763861954212, + 0.010074097663164139, + 0.004701205529272556, + 0.003936424385756254, + 0.0053554424084723, + 0.0027317553758621216, + 0.005452662706375122, + 0.003767554648220539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.033703748136758804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-speed aerial shot, mimicking the style of an Olympic live broadcast, tilts downward and pans sideways to capture a heart-pounding chase. In the frame, a legendary tengu swoops down from the summit of a jagged, snow-capped mountain, his form dwarfed yet explosive against the vast, deep blue sky. With a crimson face and a signature, unusually long nose, his expression is focused and fierce, his gaze fixed like a hawk on a distant target in the valley below. He clads himself in a pair of massive, jet-black wings, each powerful flap seeming to tear through the clear, crisp mountain air. His body, tensed into a perfectly streamlined form, impacts the ground with such incredible force and speed that even the air ripples subtly around him. The traditional Yamabushi ascetic clothing he wore danced wildly in the high-speed airflow. The whole scene, illuminated by the bright and dazzling sunlight, between the vast blue sky and the continuous snow-capped peaks, created a grand, primitive and extremely competitive atmosphere, perfectly demonstrating the absolute power and speed of this mythical creature.", + "seed": 11621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11621_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": -0.017778112733503805, + "vlm_logprob": -0.017778112733503805, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.017778112733503805, + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06253250688314438 + ], + "opposite_logprob_tokens": [ + -2.812532424926758 + ], + "logprob_reward": -0.06253250688314438, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06253250688314438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939346194267273, + 0.017496541142463684, + 0.05952336639165878, + 0.03760441765189171, + 0.18599340319633484, + 0.38572975993156433, + 0.03817114979028702, + 0.026696812361478806, + 0.022789034992456436, + 0.03206092119216919 + ], + "credit_map": [ + 0.018130574375391006, + 0.01798906922340393, + 0.015513131394982338, + 0.014759286306798458, + 0.05608614534139633, + 0.03846593573689461, + 0.014618807472288609, + 0.01837165653705597, + 0.001985718496143818, + 0.0031910999678075314, + 0.0015069848159328103, + 0.0015700282528996468, + 0.0031320396810770035, + 0.002635075245052576, + 0.00202234392054379, + 0.0014532489003613591, + 0.005130653269588947, + 0.005808783695101738, + 0.005420764442533255, + 0.0038049158174544573, + 0.014477388933300972, + 0.013662955723702908, + 0.0039965491741895676, + 0.007221348583698273, + 0.0027461256831884384, + 0.004771499428898096, + 0.003401365363970399, + 0.0018430164782330394, + 0.010590297169983387, + 0.008980113081634045, + 0.0025516552850604057, + 0.0027203415520489216, + 0.01152005698531866, + 0.011989660561084747, + 0.010081854648888111, + 0.013893239200115204, + 0.05361086502671242, + 0.050414178520441055, + 0.026886455714702606, + 0.007597071118652821, + 0.02966795302927494, + 0.05420676991343498, + 0.017549697309732437, + 0.036935821175575256, + 0.07571220397949219, + 0.06086495891213417, + 0.06823404878377914, + 0.04255825653672218, + 0.004574368242174387, + 0.004536183550953865, + 0.0024968828074634075, + 0.0035613495856523514, + 0.005673582665622234, + 0.006456425879150629, + 0.006816389039158821, + 0.004055967088788748, + 0.002544050570577383, + 0.002585828071460128, + 0.0034129659179598093, + 0.0036612667609006166, + 0.0037456220015883446, + 0.003686768701300025, + 0.0037140483036637306, + 0.003346259007230401, + 0.0030674932058900595, + 0.0017988813342526555, + 0.0028541111387312412, + 0.0025401131715625525, + 0.0024504996836185455, + 0.002665712032467127, + 0.00417220126837492, + 0.003240020480006933, + 0.00382150593213737, + 0.0034581508953124285, + 0.0037520132027566433, + 0.003909063059836626, + 0.003925623372197151, + 0.005349497776478529, + 0.002838693093508482, + 0.005006368272006512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.06253250688314438 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21957217156887054, + 0.03072229214012623, + 0.0830625593662262, + 0.10003221035003662, + 0.199436753988266, + 0.20633980631828308, + 0.05281607061624527, + 0.046156346797943115, + 0.028467074036598206, + 0.0333947129547596 + ], + "credit_map": [ + 0.016458764672279358, + 0.020304761826992035, + 0.018107319250702858, + 0.01750955916941166, + 0.06473485380411148, + 0.04643428325653076, + 0.016260866075754166, + 0.01976175606250763, + 0.0032520487438887358, + 0.003955385647714138, + 0.004072402138262987, + 0.0030245075467973948, + 0.004417043644934893, + 0.004957639146596193, + 0.0031533853616565466, + 0.003889878746122122, + 0.0059702154248952866, + 0.007585856597870588, + 0.005006931722164154, + 0.004462047014385462, + 0.023863332346081734, + 0.023590482771396637, + 0.005122110713273287, + 0.00746158417314291, + 0.010856732726097107, + 0.016900766640901566, + 0.013111517764627934, + 0.007972773164510727, + 0.013388961553573608, + 0.01602918840944767, + 0.011814148165285587, + 0.009958126582205296, + 0.01254434883594513, + 0.014230531640350819, + 0.010881132446229458, + 0.015088530257344246, + 0.057037971913814545, + 0.05225199833512306, + 0.028887556865811348, + 0.008514698594808578, + 0.017420196905732155, + 0.02471383474767208, + 0.010109808295965195, + 0.02147187478840351, + 0.03510390222072601, + 0.03293523192405701, + 0.039987094700336456, + 0.02459787204861641, + 0.006881976965814829, + 0.007768961135298014, + 0.005203314125537872, + 0.007379868067800999, + 0.007299905642867088, + 0.0069305733777582645, + 0.005725383292883635, + 0.005626086611300707, + 0.006883713882416487, + 0.005126469302922487, + 0.005759142339229584, + 0.004326728638261557, + 0.006158050149679184, + 0.009134367108345032, + 0.0038178071845322847, + 0.0049500674940645695, + 0.003252514638006687, + 0.0023215683177113533, + 0.0030694096349179745, + 0.0030879382975399494, + 0.004272080026566982, + 0.004949103109538555, + 0.0037538763135671616, + 0.003760585328564048, + 0.003316039452329278, + 0.0031257430091500282, + 0.0038139407988637686, + 0.00353897619061172, + 0.005511254072189331, + 0.005109211429953575, + 0.004699158947914839, + 0.004280388820916414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0067449514754116535 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003412379592191428 + ], + "opposite_logprob_tokens": [ + -8.000341415405273 + ], + "logprob_reward": -0.0003412379592191428, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0003412379592191428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25238096714019775, + 0.07421983778476715, + 0.056031398475170135, + 0.055850304663181305, + 0.11768888682126999, + 0.2731386721134186, + 0.024206427857279778, + 0.05293255299329758, + 0.032856687903404236, + 0.06069432199001312 + ], + "credit_map": [ + 0.027519937604665756, + 0.02197588048875332, + 0.014630194753408432, + 0.00885196402668953, + 0.09205152094364166, + 0.06273966282606125, + 0.009539863094687462, + 0.015071949921548367, + 0.009605882689356804, + 0.01368793286383152, + 0.00836089625954628, + 0.007226389832794666, + 0.008090149611234665, + 0.01021352969110012, + 0.009186149574816227, + 0.007848906330764294, + 0.005742282141000032, + 0.007879899814724922, + 0.006079448852688074, + 0.004345490597188473, + 0.010696965269744396, + 0.008944811299443245, + 0.0046063195914030075, + 0.007736182771623135, + 0.005824967287480831, + 0.006421769503504038, + 0.007877983152866364, + 0.0033406077418476343, + 0.00932076945900917, + 0.008870591409504414, + 0.004713401198387146, + 0.00948021188378334, + 0.007838694378733635, + 0.01156273391097784, + 0.010630287230014801, + 0.010564073920249939, + 0.02607974410057068, + 0.03053208254277706, + 0.015319365076720715, + 0.005161906126886606, + 0.02169492095708847, + 0.03949466347694397, + 0.013061189092695713, + 0.026740662753582, + 0.050395309925079346, + 0.04204335808753967, + 0.04761886969208717, + 0.03208967670798302, + 0.0025206629652529955, + 0.0030724620446562767, + 0.001935593318194151, + 0.0026830448769032955, + 0.002931246068328619, + 0.003545655868947506, + 0.0034766874741762877, + 0.004041075706481934, + 0.012682311236858368, + 0.0025543272495269775, + 0.003603338962420821, + 0.003382609924301505, + 0.007515243254601955, + 0.01764121651649475, + 0.0024306117556989193, + 0.003122894559055567, + 0.00235308357514441, + 0.001968718832358718, + 0.003115243511274457, + 0.002717040479183197, + 0.005677387118339539, + 0.0064707533456385136, + 0.00579166179522872, + 0.004762798547744751, + 0.006093147210776806, + 0.007770640775561333, + 0.0073279524222016335, + 0.008382675237953663, + 0.009066714905202389, + 0.009305491112172604, + 0.005066210869699717, + 0.007681489922106266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0003412379592191428 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731334142386913 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731334142386913, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731334142386913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17773912847042084, + 0.0615849606692791, + 0.06680712103843689, + 0.08797933161258698, + 0.11299438774585724, + 0.14025653898715973, + 0.06802678108215332, + 0.037469491362571716, + 0.06625193357467651, + 0.18089035153388977 + ], + "credit_map": [ + 0.01726432517170906, + 0.019969159737229347, + 0.01929883472621441, + 0.01170534547418356, + 0.0322001688182354, + 0.03757455199956894, + 0.014549234881997108, + 0.025177491828799248, + 0.008061045780777931, + 0.011078491806983948, + 0.007936853915452957, + 0.004937571007758379, + 0.00789697002619505, + 0.009340779855847359, + 0.007020859979093075, + 0.005312385503202677, + 0.005430791527032852, + 0.007821008563041687, + 0.006613448262214661, + 0.004195793066173792, + 0.01563819870352745, + 0.014909547753632069, + 0.0047179763205349445, + 0.0074803526513278484, + 0.011036287061870098, + 0.010893497616052628, + 0.012905340641736984, + 0.008010799065232277, + 0.009877908043563366, + 0.012365804985165596, + 0.013684939593076706, + 0.009204743430018425, + 0.010277904570102692, + 0.014734393917024136, + 0.01365750189870596, + 0.01597341150045395, + 0.010707964189350605, + 0.01309320330619812, + 0.0261925607919693, + 0.008357434533536434, + 0.009283111430704594, + 0.025493690744042397, + 0.005809359718114138, + 0.01189344096928835, + 0.03491605073213577, + 0.01971912384033203, + 0.017421187832951546, + 0.015720561146736145, + 0.0060753184370696545, + 0.0100026223808527, + 0.005669678095728159, + 0.00957406684756279, + 0.009114331565797329, + 0.00962058361619711, + 0.009639784693717957, + 0.008330389857292175, + 0.0047953068278729916, + 0.004115193150937557, + 0.004752056673169136, + 0.0045425477437675, + 0.0040824683383107185, + 0.005724536720663309, + 0.0052788518369197845, + 0.0041785272769629955, + 0.0081363869830966, + 0.004558858927339315, + 0.008336188271641731, + 0.006560472305864096, + 0.008848646655678749, + 0.007287764456123114, + 0.0104931415989995, + 0.012030471116304398, + 0.016261519864201546, + 0.021822135895490646, + 0.02398388646543026, + 0.020612062886357307, + 0.027180049568414688, + 0.022850412875413895, + 0.024023201316595078, + 0.024157077074050903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.006731334142386913 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254053320735693 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254053320735693, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254053320735693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4638519883155823, + 0.04873919114470482, + 0.04959537833929062, + 0.059192121028900146, + 0.04771370440721512, + 0.1463531106710434, + 0.050123002380132675, + 0.03803825378417969, + 0.038742806762456894, + 0.057650547474622726 + ], + "credit_map": [ + 0.03616633638739586, + 0.04997909441590309, + 0.03969203308224678, + 0.03330393135547638, + 0.11353455483913422, + 0.10531359165906906, + 0.036698222160339355, + 0.049164123833179474, + 0.005532577633857727, + 0.007695022039115429, + 0.006625946145504713, + 0.003301728516817093, + 0.008435850031673908, + 0.007238078396767378, + 0.004816312808543444, + 0.005093665793538094, + 0.00632801279425621, + 0.005917344242334366, + 0.0069238850846886635, + 0.004285592120140791, + 0.008389031514525414, + 0.00869743898510933, + 0.003909657243639231, + 0.005144407972693443, + 0.006069946102797985, + 0.006691889371722937, + 0.005957477726042271, + 0.0047180247493088245, + 0.01106902863830328, + 0.011649906635284424, + 0.008219441398978233, + 0.004816395230591297, + 0.00460847606882453, + 0.006805419921875, + 0.004138777032494545, + 0.005339584778994322, + 0.005461586639285088, + 0.0065986295230686665, + 0.010887830518186092, + 0.0038733896799385548, + 0.007582615129649639, + 0.03531274199485779, + 0.004836540203541517, + 0.006878600921481848, + 0.048861946910619736, + 0.01960585080087185, + 0.013249268755316734, + 0.010025519877672195, + 0.004677161108702421, + 0.006787147372961044, + 0.0035963039845228195, + 0.005719437263906002, + 0.005803563632071018, + 0.009636457078158855, + 0.009024213068187237, + 0.00487870816141367, + 0.005902208853513002, + 0.0036733488086611032, + 0.0041323015466332436, + 0.003996620886027813, + 0.003855409799143672, + 0.00785468053072691, + 0.003822427010163665, + 0.004801247268915176, + 0.004883151967078447, + 0.0036249326076358557, + 0.004353000316768885, + 0.004109631292521954, + 0.0047571673057973385, + 0.005423558875918388, + 0.004817846231162548, + 0.006773508153855801, + 0.005978392902761698, + 0.0050709606148302555, + 0.007752918638288975, + 0.0076185064390301704, + 0.005734169390052557, + 0.005996504798531532, + 0.007926828227937222, + 0.011572254821658134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.01254053320735693 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-speed aerial shot, mimicking the style of an Olympic live broadcast, tilts downward and pans sideways to capture a heart-pounding chase. In the frame, a legendary tengu swoops down from the summit of a jagged, snow-capped mountain, his form dwarfed yet explosive against the vast, deep blue sky. With a crimson face and a signature, unusually long nose, his expression is focused and fierce, his gaze fixed like a hawk on a distant target in the valley below. He clads himself in a pair of massive, jet-black wings, each powerful flap seeming to tear through the clear, crisp mountain air. His body, tensed into a perfectly streamlined form, impacts the ground with such incredible force and speed that even the air ripples subtly around him. The traditional Yamabushi ascetic clothing he wore danced wildly in the high-speed airflow. The whole scene, illuminated by the bright and dazzling sunlight, between the vast blue sky and the continuous snow-capped peaks, created a grand, primitive and extremely competitive atmosphere, perfectly demonstrating the absolute power and speed of this mythical creature.", + "seed": 11622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11622_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": -0.01853956896811724, + "vlm_logprob": -0.01853956896811724, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01853956896811724, + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056112103164196014 + ], + "opposite_logprob_tokens": [ + -2.931112051010132 + ], + "logprob_reward": -0.056112103164196014, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.056112103164196014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04790440946817398, + 0.02689354121685028, + 0.03925132378935814, + 0.27240580320358276, + 0.5293994545936584, + 0.02666526474058628, + 0.010460716672241688, + 0.01737876981496811, + 0.021432844921946526, + 0.00820781197398901 + ], + "credit_map": [ + 0.00569232227280736, + 0.00608349684625864, + 0.006092061288654804, + 0.004292943049222231, + 0.008030764758586884, + 0.007958602160215378, + 0.004968257620930672, + 0.004785963799804449, + 0.0033739113714545965, + 0.00305243581533432, + 0.0030562535393983126, + 0.004327491391450167, + 0.0039657000452280045, + 0.0026924151461571455, + 0.003922034986317158, + 0.0025033005513250828, + 0.003194080665707588, + 0.00882614403963089, + 0.002376008778810501, + 0.004135393537580967, + 0.010057044215500355, + 0.004590862430632114, + 0.0031758516561239958, + 0.002895940328016877, + 0.006960997357964516, + 0.02820662222802639, + 0.0031676236540079117, + 0.004096624441444874, + 0.10083047300577164, + 0.12274289131164551, + 0.0025717734824866056, + 0.003828788874670863, + 0.017721980810165405, + 0.029121818020939827, + 0.012710818089544773, + 0.042668573558330536, + 0.19897617399692535, + 0.19592101871967316, + 0.01879291795194149, + 0.013486149720847607, + 0.002512986771762371, + 0.005661095026880503, + 0.0022057583555579185, + 0.004143607337027788, + 0.00311892107129097, + 0.003472036449238658, + 0.0030121305026113987, + 0.0025387282948940992, + 0.0013400801690295339, + 0.001419326290488243, + 0.001834618509747088, + 0.0007897301111370325, + 0.0018243343802168965, + 0.0013240550179034472, + 0.0009291196474805474, + 0.0009994521969929338, + 0.002224144758656621, + 0.0023068785667419434, + 0.0015833029756322503, + 0.002155224559828639, + 0.0019078829791396856, + 0.002357109682634473, + 0.0020002909004688263, + 0.0028439362067729235, + 0.001658739405684173, + 0.0035588329192250967, + 0.003099884372204542, + 0.0026276663411408663, + 0.0027723636012524366, + 0.0025799712166190147, + 0.002751399762928486, + 0.002383986720815301, + 0.0006036304985173047, + 0.0014362597139552236, + 0.0007221493287943304, + 0.0005937724490649998, + 0.0018329888116568327, + 0.0014857820933684707, + 0.0006907404749654233, + 0.0008424887200817466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.056112103164196014 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639833260327578 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639833260327578, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639833260327578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24522539973258972, + 0.039750441908836365, + 0.1430756002664566, + 0.09795387834310532, + 0.343645304441452, + 0.03579883277416229, + 0.018424399197101593, + 0.03935450688004494, + 0.022740500047802925, + 0.014031145721673965 + ], + "credit_map": [ + 0.02365368977189064, + 0.04159456863999367, + 0.02572258748114109, + 0.02147727645933628, + 0.048692330718040466, + 0.040689848363399506, + 0.025395771488547325, + 0.01799933612346649, + 0.003976551350206137, + 0.0033569561783224344, + 0.0039613558910787106, + 0.006261959206312895, + 0.00604601763188839, + 0.004272105637937784, + 0.006505670957267284, + 0.005369824357330799, + 0.011765257455408573, + 0.03560178726911545, + 0.009946734644472599, + 0.014516607858240604, + 0.027537938207387924, + 0.020195357501506805, + 0.013442954048514366, + 0.010068965144455433, + 0.005341588519513607, + 0.015736842527985573, + 0.005211250856518745, + 0.005027828272432089, + 0.03531105816364288, + 0.021611209958791733, + 0.0038048080168664455, + 0.005909298080950975, + 0.010984729044139385, + 0.0178506001830101, + 0.007711120881140232, + 0.024837398901581764, + 0.13380849361419678, + 0.12987174093723297, + 0.010197053663432598, + 0.008384162560105324, + 0.0034452599938958883, + 0.005873116664588451, + 0.002792511135339737, + 0.004785280209034681, + 0.006361052859574556, + 0.00641141552478075, + 0.003089133882895112, + 0.0030410641338676214, + 0.0034633087925612926, + 0.0018811464542523026, + 0.0023623357992619276, + 0.0020116136875003576, + 0.0022009829990565777, + 0.0018768442096188664, + 0.0016799397999420762, + 0.002948226872831583, + 0.005184968933463097, + 0.003929538652300835, + 0.004592529032379389, + 0.003702538786455989, + 0.007109233643859625, + 0.006520912516862154, + 0.0035419631749391556, + 0.00477282190695405, + 0.001659109373576939, + 0.003864787984639406, + 0.002826183568686247, + 0.0019037830643355846, + 0.004185054451227188, + 0.0036737839691340923, + 0.0027274456806480885, + 0.0019003519555553794, + 0.0010496482718735933, + 0.0026390263810753822, + 0.0010958736529573798, + 0.0011275499127805233, + 0.0031238768715411425, + 0.0025088987313210964, + 0.0016506797401234508, + 0.0008355934405699372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.007639833260327578 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005601267330348492 + ], + "opposite_logprob_tokens": [ + -7.500560283660889 + ], + "logprob_reward": -0.0005601267330348492, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005601267330348492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15644563734531403, + 0.03139748051762581, + 0.046728380024433136, + 0.09276711195707321, + 0.5353727340698242, + 0.021717682480812073, + 0.044272828847169876, + 0.038527801632881165, + 0.019876612350344658, + 0.012893695384263992 + ], + "credit_map": [ + 0.021765850484371185, + 0.020645586773753166, + 0.02197912521660328, + 0.014402152970433235, + 0.01659179851412773, + 0.017604093998670578, + 0.01786934770643711, + 0.025587690994143486, + 0.004269878845661879, + 0.003312372136861086, + 0.003511464688926935, + 0.0035878163762390614, + 0.004806290380656719, + 0.002775880042463541, + 0.0037418329156935215, + 0.00539194792509079, + 0.004861658439040184, + 0.008552116341888905, + 0.003067862009629607, + 0.004489515442401171, + 0.011569451540708542, + 0.0071534086018800735, + 0.0028486468363553286, + 0.004185722209513187, + 0.0047469413839280605, + 0.014754679054021835, + 0.0027462271973490715, + 0.002253961283713579, + 0.025508850812911987, + 0.038371093571186066, + 0.0017236535204574466, + 0.002661709673702717, + 0.00994771346449852, + 0.019058361649513245, + 0.008579012006521225, + 0.00716610811650753, + 0.2467445284128189, + 0.2310446798801422, + 0.00593846570700407, + 0.006893815938383341, + 0.002458092523738742, + 0.002318906830623746, + 0.0026790881529450417, + 0.0018576536094769835, + 0.0035021693911403418, + 0.004058630205690861, + 0.0019808157812803984, + 0.002862327266484499, + 0.007026528939604759, + 0.005036365706473589, + 0.004682329948991537, + 0.006039012689143419, + 0.005278967786580324, + 0.0044437190517783165, + 0.004636987112462521, + 0.00712891761213541, + 0.004423195030540228, + 0.0034154076129198074, + 0.005614206660538912, + 0.0039007391314953566, + 0.004842520225793123, + 0.005480165593326092, + 0.002720544347539544, + 0.008131024427711964, + 0.0013091261498630047, + 0.0025883791968226433, + 0.0018120704917237163, + 0.0016364033799618483, + 0.004866369534283876, + 0.0038709831424057484, + 0.0022863822523504496, + 0.0015068999491631985, + 0.001485647982917726, + 0.002368250861763954, + 0.001371113583445549, + 0.0012431871145963669, + 0.0018557871226221323, + 0.0020197387784719467, + 0.0013799261068925261, + 0.0011700444156304002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0005601267330348492 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187073335051537 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014187073335051537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187073335051537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16955547034740448, + 0.0692654699087143, + 0.07632392644882202, + 0.1699354499578476, + 0.3242432177066803, + 0.06588713079690933, + 0.0428754985332489, + 0.03377918899059296, + 0.03409738093614578, + 0.014037322252988815 + ], + "credit_map": [ + 0.017932415008544922, + 0.03167446330189705, + 0.016484659165143967, + 0.015181438066065311, + 0.03306873142719269, + 0.02551145665347576, + 0.016739463433623314, + 0.012962832115590572, + 0.009098989889025688, + 0.007764514070004225, + 0.007100837305188179, + 0.009650477208197117, + 0.011795050464570522, + 0.009492523036897182, + 0.008221653290092945, + 0.006141420919448137, + 0.009043436497449875, + 0.01546062808483839, + 0.004938553553074598, + 0.005770833231508732, + 0.012801773846149445, + 0.013141882605850697, + 0.008854174986481667, + 0.006312637124210596, + 0.00464095501229167, + 0.019070811569690704, + 0.002963038394227624, + 0.003833303228020668, + 0.05869463086128235, + 0.07464981824159622, + 0.002300501801073551, + 0.0037823617458343506, + 0.007741960231214762, + 0.010643510147929192, + 0.00661878939718008, + 0.0075400276109576225, + 0.14427246153354645, + 0.13649529218673706, + 0.006417258642613888, + 0.004513890482485294, + 0.005912396591156721, + 0.011082123965024948, + 0.0041385069489479065, + 0.00501422630622983, + 0.01386499498039484, + 0.01303163357079029, + 0.005845090840011835, + 0.006998151540756226, + 0.005005777347832918, + 0.006305965594947338, + 0.004128233529627323, + 0.006890658289194107, + 0.005878741387277842, + 0.004868554417043924, + 0.005022153258323669, + 0.0047754100523889065, + 0.0035310694947838783, + 0.004284879192709923, + 0.003485947148874402, + 0.0034533932339400053, + 0.005119154695421457, + 0.00477873208001256, + 0.003277632873505354, + 0.005848372355103493, + 0.0038626280147582293, + 0.004000439774245024, + 0.004766055382788181, + 0.0030766823329031467, + 0.006614189129322767, + 0.0054903230629861355, + 0.002892777556553483, + 0.0033942812588065863, + 0.0014197443379089236, + 0.0022363204043358564, + 0.002066703513264656, + 0.0017450778977945447, + 0.001805005711503327, + 0.0017420968506485224, + 0.0016370874363929033, + 0.0013852835400030017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.014187073335051537 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198708347976208 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198708347976208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198708347976208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13498766720294952, + 0.10000776499509811, + 0.10293801873922348, + 0.2606455385684967, + 0.14722314476966858, + 0.06030561774969101, + 0.04722537472844124, + 0.044258248060941696, + 0.06147260591387749, + 0.04093599319458008 + ], + "credit_map": [ + 0.015411624684929848, + 0.02403913252055645, + 0.013823688961565495, + 0.010042711161077023, + 0.025900062173604965, + 0.01854962669312954, + 0.014605702832341194, + 0.012615127488970757, + 0.014082135632634163, + 0.011953400447964668, + 0.008000937290489674, + 0.012187921442091465, + 0.02036801539361477, + 0.013104242272675037, + 0.011325943283736706, + 0.00898517295718193, + 0.010843012481927872, + 0.012971713207662106, + 0.00725000211969018, + 0.012597598135471344, + 0.017827097326517105, + 0.018426934257149696, + 0.013947204686701298, + 0.009074464440345764, + 0.005900911521166563, + 0.01838941127061844, + 0.003853960894048214, + 0.004757492803037167, + 0.1059069111943245, + 0.11271707713603973, + 0.0035310997627675533, + 0.005588680040091276, + 0.010686916299164295, + 0.01947021670639515, + 0.007275691721588373, + 0.019602743908762932, + 0.0374322235584259, + 0.032305147498846054, + 0.011417175643146038, + 0.009033037349581718, + 0.005496721249073744, + 0.007299952674657106, + 0.005400089081376791, + 0.004705250263214111, + 0.012040967121720314, + 0.01308742631226778, + 0.005504243541508913, + 0.006770973093807697, + 0.006351812742650509, + 0.009104390628635883, + 0.004690444562584162, + 0.004793265834450722, + 0.006771007552742958, + 0.005049927160143852, + 0.0051546660251915455, + 0.005309869069606066, + 0.0036287568509578705, + 0.006533490493893623, + 0.0033349760342389345, + 0.006887002848088741, + 0.008153197355568409, + 0.0071585215628147125, + 0.004121247213333845, + 0.0044410633854568005, + 0.007234289310872555, + 0.006803716067224741, + 0.008980536833405495, + 0.005635002627968788, + 0.01151945348829031, + 0.010035046376287937, + 0.003934203181415796, + 0.007330362685024738, + 0.0038907243870198727, + 0.007608475629240274, + 0.006101721432060003, + 0.004919096827507019, + 0.0056417821906507015, + 0.005033775698393583, + 0.003977078478783369, + 0.0037633441388607025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.014198708347976208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, high-speed aerial shot, mimicking the style of an Olympic live broadcast, tilts downward and pans sideways to capture a heart-pounding chase. In the frame, a legendary tengu swoops down from the summit of a jagged, snow-capped mountain, his form dwarfed yet explosive against the vast, deep blue sky. With a crimson face and a signature, unusually long nose, his expression is focused and fierce, his gaze fixed like a hawk on a distant target in the valley below. He clads himself in a pair of massive, jet-black wings, each powerful flap seeming to tear through the clear, crisp mountain air. His body, tensed into a perfectly streamlined form, impacts the ground with such incredible force and speed that even the air ripples subtly around him. The traditional Yamabushi ascetic clothing he wore danced wildly in the high-speed airflow. The whole scene, illuminated by the bright and dazzling sunlight, between the vast blue sky and the continuous snow-capped peaks, created a grand, primitive and extremely competitive atmosphere, perfectly demonstrating the absolute power and speed of this mythical creature.", + "seed": 11623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_7/index_6082_seed_11623_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": -0.02290590035263449, + "vlm_logprob": -0.02290590035263449, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02290590035263449, + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055391211062669754 + ], + "opposite_logprob_tokens": [ + -2.930391311645508 + ], + "logprob_reward": -0.055391211062669754, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055391211062669754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10747437924146652, + 0.05299996957182884, + 0.06962309777736664, + 0.02505670301616192, + 0.03653130680322647, + 0.06047798693180084, + 0.04406868666410446, + 0.0358569510281086, + 0.334179162979126, + 0.2337317317724228 + ], + "credit_map": [ + 0.007153115700930357, + 0.010101721622049809, + 0.009684937074780464, + 0.006292314734309912, + 0.011236073449254036, + 0.03415204957127571, + 0.00630504684522748, + 0.022549116984009743, + 0.006165151949971914, + 0.007498721592128277, + 0.013040012679994106, + 0.00594426179304719, + 0.007298261858522892, + 0.00628329161554575, + 0.003996703308075666, + 0.0027735689654946327, + 0.009158745408058167, + 0.006638773716986179, + 0.015481174923479557, + 0.00658615306019783, + 0.008549870923161507, + 0.007654170505702496, + 0.008748134598135948, + 0.006806074641644955, + 0.0029941771645098925, + 0.0038470060098916292, + 0.003594394540414214, + 0.002103302162140608, + 0.0046533988788723946, + 0.0025755190290510654, + 0.0035339065361768007, + 0.001754998927935958, + 0.0032336001750081778, + 0.0046156723983585835, + 0.00557379424571991, + 0.004479711409658194, + 0.005147913470864296, + 0.004647525027394295, + 0.004280876833945513, + 0.004552212078124285, + 0.007380088325589895, + 0.007035326678305864, + 0.008616727776825428, + 0.007933923974633217, + 0.009420778602361679, + 0.008252285420894623, + 0.005242165178060532, + 0.006596691440790892, + 0.004439277574419975, + 0.005948603618890047, + 0.005007628351449966, + 0.003526900662109256, + 0.006786775309592485, + 0.006340977270156145, + 0.0051681422628462315, + 0.0068503799848258495, + 0.003190449671819806, + 0.0036413162015378475, + 0.00751170190051198, + 0.002295765560120344, + 0.006519034039229155, + 0.005275064613670111, + 0.0036892383359372616, + 0.00373437674716115, + 0.029037870466709137, + 0.06098169460892677, + 0.030015435069799423, + 0.032594237476587296, + 0.055723998695611954, + 0.0416993610560894, + 0.04455108940601349, + 0.03957551345229149, + 0.031245112419128418, + 0.05156891047954559, + 0.020693307742476463, + 0.04106391593813896, + 0.021415114402770996, + 0.022968661040067673, + 0.021091768518090248, + 0.02368495613336563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.055391211062669754 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254853792488575 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.01254853792488575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254853792488575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10074777901172638, + 0.05302047356963158, + 0.10340192168951035, + 0.0608164481818676, + 0.07158021628856659, + 0.0897606611251831, + 0.06545476615428925, + 0.05997832491993904, + 0.22539959847927094, + 0.16983981430530548 + ], + "credit_map": [ + 0.009123032912611961, + 0.015549683012068272, + 0.013936125673353672, + 0.007672241888940334, + 0.011107569560408592, + 0.02552730217576027, + 0.007080599199980497, + 0.010751232504844666, + 0.006414565723389387, + 0.0070908269844949245, + 0.008675584569573402, + 0.006072500254958868, + 0.007429236546158791, + 0.007661344017833471, + 0.004447086248546839, + 0.005229331087321043, + 0.013983313925564289, + 0.012180496007204056, + 0.022951480001211166, + 0.009502861648797989, + 0.013030150905251503, + 0.011308656074106693, + 0.011634109541773796, + 0.008810858242213726, + 0.0070168571546673775, + 0.007193912751972675, + 0.00754427770152688, + 0.006115032825618982, + 0.00965258851647377, + 0.007302724290639162, + 0.00922743696719408, + 0.006763618439435959, + 0.005191292148083448, + 0.008016715757548809, + 0.008076242171227932, + 0.01867765188217163, + 0.012148568406701088, + 0.004878037143498659, + 0.009394004940986633, + 0.005197702907025814, + 0.010096350684762001, + 0.013794684782624245, + 0.015715351328253746, + 0.010554763488471508, + 0.009635982103645802, + 0.009054882451891899, + 0.009847236797213554, + 0.011061408556997776, + 0.0061931973323225975, + 0.011209328658878803, + 0.005254206247627735, + 0.005621802993118763, + 0.010381796397268772, + 0.009407512843608856, + 0.009128729812800884, + 0.008258186280727386, + 0.005960014648735523, + 0.006929140537977219, + 0.009999928995966911, + 0.00471863616257906, + 0.010340472683310509, + 0.007402317598462105, + 0.007892772555351257, + 0.006735046394169331, + 0.016632037237286568, + 0.03520679101347923, + 0.02047417126595974, + 0.02001500315964222, + 0.03656863421201706, + 0.013849337585270405, + 0.05743476748466492, + 0.025218836963176727, + 0.025865545496344566, + 0.02147773653268814, + 0.03207441046833992, + 0.021014804020524025, + 0.018452834337949753, + 0.017842059955000877, + 0.01629740558564663, + 0.016815004870295525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.01254853792488575 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007167869480326772 + ], + "opposite_logprob_tokens": [ + -7.250716686248779 + ], + "logprob_reward": -0.0007167869480326772, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007167869480326772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1078668013215065, + 0.16347190737724304, + 0.0859001949429512, + 0.2659863829612732, + 0.06257642060518265, + 0.049729835242033005, + 0.05571381747722626, + 0.03638837859034538, + 0.052102118730545044, + 0.12026414275169373 + ], + "credit_map": [ + 0.00671392073854804, + 0.008692191913723946, + 0.011625797487795353, + 0.006354517303407192, + 0.011005721054971218, + 0.03618001937866211, + 0.007765518967062235, + 0.019529107958078384, + 0.022693805396556854, + 0.019998271018266678, + 0.021487919613718987, + 0.017199281603097916, + 0.021561693400144577, + 0.026217058300971985, + 0.01708809658885002, + 0.01722577214241028, + 0.013118764385581017, + 0.008352606557309628, + 0.012387615628540516, + 0.008444700390100479, + 0.010340478271245956, + 0.010019179433584213, + 0.013912178575992584, + 0.009324668906629086, + 0.030063889920711517, + 0.021764183416962624, + 0.04442762956023216, + 0.02797815389931202, + 0.02825329825282097, + 0.03749554604291916, + 0.03773967921733856, + 0.038263995200395584, + 0.007009935565292835, + 0.007440662942826748, + 0.0077804806642234325, + 0.007044149562716484, + 0.00850977934896946, + 0.007419577334076166, + 0.008938330225646496, + 0.008433511480689049, + 0.005459070671349764, + 0.006319169886410236, + 0.007085128221660852, + 0.005963244941085577, + 0.009236780926585197, + 0.005590518936514854, + 0.004490343388170004, + 0.005585579201579094, + 0.009193896315991879, + 0.004948751535266638, + 0.00965501181781292, + 0.004578273743391037, + 0.006190196610987186, + 0.007303361315280199, + 0.003762667765840888, + 0.010081663727760315, + 0.0046557714231312275, + 0.0045366655103862286, + 0.004982981830835342, + 0.0027052327059209347, + 0.004879362881183624, + 0.004978872369974852, + 0.005389030557125807, + 0.004260456655174494, + 0.005384781863540411, + 0.009784525260329247, + 0.004934622906148434, + 0.004266636446118355, + 0.009868742898106575, + 0.005320825148373842, + 0.007924729958176613, + 0.004617250990122557, + 0.017029594630002975, + 0.020525963976979256, + 0.027278097346425056, + 0.01888432167470455, + 0.009226159192621708, + 0.009140729904174805, + 0.010026093572378159, + 0.008153174072504044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.0007167869480326772 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606203243136406 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606203243136406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606203243136406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17625273764133453, + 0.0956084206700325, + 0.15299087762832642, + 0.13468928635120392, + 0.07567591965198517, + 0.06729590892791748, + 0.04709947854280472, + 0.06981318444013596, + 0.10438035428524017, + 0.07619387656450272 + ], + "credit_map": [ + 0.011937550269067287, + 0.008972330018877983, + 0.02295183390378952, + 0.007542781997472048, + 0.013627919368445873, + 0.06700222939252853, + 0.008464503102004528, + 0.03575356304645538, + 0.012200904078781605, + 0.01038117241114378, + 0.01891176402568817, + 0.011353341862559319, + 0.008327481336891651, + 0.012090716511011124, + 0.008832719177007675, + 0.013510310091078281, + 0.01830625720322132, + 0.022122198715806007, + 0.0199615266174078, + 0.009730068035423756, + 0.029971178621053696, + 0.015928607434034348, + 0.02289622649550438, + 0.014074794016778469, + 0.013940331526100636, + 0.01215311698615551, + 0.022350234910845757, + 0.013512634672224522, + 0.014893226325511932, + 0.018954835832118988, + 0.018240446224808693, + 0.020644456148147583, + 0.006933862809091806, + 0.009128500707447529, + 0.010031843557953835, + 0.015049994923174381, + 0.007994208484888077, + 0.007821482606232166, + 0.011628055945038795, + 0.00708796177059412, + 0.007097519468516111, + 0.009124006144702435, + 0.008426571264863014, + 0.01158479880541563, + 0.010768471285700798, + 0.008740969933569431, + 0.0062035806477069855, + 0.005349981132894754, + 0.004254556261003017, + 0.00515070790424943, + 0.008751105517148972, + 0.00417162524536252, + 0.005581726785749197, + 0.005773757118731737, + 0.005590763408690691, + 0.007825233973562717, + 0.008175252005457878, + 0.008285574615001678, + 0.014988374896347523, + 0.004117341712117195, + 0.006981467828154564, + 0.009445992298424244, + 0.008365476503968239, + 0.009453697130084038, + 0.008824649266898632, + 0.01916000247001648, + 0.010706454515457153, + 0.0070440624840557575, + 0.018458249047398567, + 0.009620890952646732, + 0.018577752634882927, + 0.011988291516900063, + 0.01156008429825306, + 0.011894473806023598, + 0.01314975880086422, + 0.010813649743795395, + 0.007184676826000214, + 0.007784747518599033, + 0.007018729578703642, + 0.00678775180131197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.01606203243136406 + }, + { + "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 tengu with a crimson face and an unusually long nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tengu wear massive jet-black wings and traditional Yamabushi ascetic clothing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a jagged, snow-capped mountain with a vast deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tengu swoop down from the summit and impact the ground with high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt downward and pan sideways while the scene is illuminated by bright sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810933396220207 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810933396220207, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810933396220207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18474741280078888, + 0.2167423814535141, + 0.08911141753196716, + 0.047809988260269165, + 0.08381491154432297, + 0.04389530047774315, + 0.026435548439621925, + 0.08082639425992966, + 0.13595741987228394, + 0.09065916389226913 + ], + "credit_map": [ + 0.016197890043258667, + 0.025348175317049026, + 0.027918817475438118, + 0.012558282352983952, + 0.020575344562530518, + 0.05076400935649872, + 0.01121916901320219, + 0.02016572840511799, + 0.013095015659928322, + 0.013300449587404728, + 0.07740958780050278, + 0.016595713794231415, + 0.01531582698225975, + 0.05827469378709793, + 0.009135936386883259, + 0.013615146279335022, + 0.009379621595144272, + 0.010676922276616096, + 0.010882790200412273, + 0.007749440148472786, + 0.018691109493374825, + 0.010754869319498539, + 0.012045095674693584, + 0.008931577205657959, + 0.006903399713337421, + 0.007225397974252701, + 0.005957619287073612, + 0.004925158340483904, + 0.007543947082012892, + 0.005410040263086557, + 0.005774420220404863, + 0.004070004913955927, + 0.00676782988011837, + 0.010094099678099155, + 0.010901816189289093, + 0.015802592039108276, + 0.013270427472889423, + 0.008397331461310387, + 0.010159376077353954, + 0.008421439677476883, + 0.0042182630859315395, + 0.004975670017302036, + 0.006391206290572882, + 0.0069228895008563995, + 0.007964099757373333, + 0.005658079870045185, + 0.004056643694639206, + 0.0037084496580064297, + 0.002174179768189788, + 0.0034876475110650063, + 0.0026461509987711906, + 0.0020979614928364754, + 0.004257943481206894, + 0.003816520795226097, + 0.004032622091472149, + 0.003922524396330118, + 0.012338714674115181, + 0.007465525064617395, + 0.02022223174571991, + 0.004378252197057009, + 0.007099767215549946, + 0.012753044255077839, + 0.009663484990596771, + 0.006905375514179468, + 0.01305034477263689, + 0.022503746673464775, + 0.01798980124294758, + 0.013044517487287521, + 0.017168143764138222, + 0.015349374152719975, + 0.018654318526387215, + 0.018197180703282356, + 0.014418784528970718, + 0.012855170294642448, + 0.019254401326179504, + 0.011442773044109344, + 0.006578143686056137, + 0.008485110476613045, + 0.009008673951029778, + 0.00861610658466816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.029810933396220207 + } + ], + "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/0000060/group_8/index_tensor([7413, 7413])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_tensor([7413, 7413])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b808b7a3877bcf825264b14a6e242f6c710ef93b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_tensor([7413, 7413])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot, rendered in a vibrant animation style with bold silhouettes and highly saturated colors, follows a classic American cruiser trike around a 3,000-meter steeplechase course. The trike features a gleaming cherry-red frame, wide, swept-back chrome handlebars, and a large, padded black saddle. Iridescent satin streamers tied to the handlebar grips and rear axle flutter wildly in its wake. The scene is set against a brilliant blue sky, where a large, stylized cartoon sun with a smiling face radiates a warm, cheerful glow. As the trike approaches a water obstacle, the camera remains low, emphasizing the impending impact. The trike dramatically splashes into the vibrant, azure water of the Danger Pit, its large front wheel impacting the surface and launching a perfect, shimmering arc of water into the air. Each cartoon-like droplet in the spray catches the sunlight, shimmering with magical intensity before falling back down, creating a moment of slow-motion, joyful wonder amidst the dynamic race.", + "seed": 27925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_7413_seed_27925_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": -0.011938122101128101, + "vlm_logprob": -0.011938122101128101, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.011938122101128101, + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025766756385564804 + ], + "opposite_logprob_tokens": [ + -3.7757668495178223 + ], + "logprob_reward": -0.025766756385564804, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025766756385564804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19610127806663513, + 0.09641804546117783, + 0.10161009430885315, + 0.0646258145570755, + 0.061769332736730576, + 0.1312515139579773, + 0.04941326379776001, + 0.15914665162563324, + 0.08264656364917755, + 0.05701737478375435 + ], + "credit_map": [ + 0.023992221802473068, + 0.048941437155008316, + 0.023978503420948982, + 0.023964377120137215, + 0.02162780612707138, + 0.017452513799071312, + 0.019951310008764267, + 0.01619313284754753, + 0.009945188648998737, + 0.020342163741588593, + 0.00908535998314619, + 0.012013017199933529, + 0.018087998032569885, + 0.012474080547690392, + 0.00680968351662159, + 0.007660568691790104, + 0.016602929681539536, + 0.01000588946044445, + 0.008585145696997643, + 0.014639468863606453, + 0.017434172332286835, + 0.016851142048835754, + 0.008582727052271366, + 0.008908629417419434, + 0.0072919144295156, + 0.009155405685305595, + 0.008803177624940872, + 0.009736456908285618, + 0.012493434362113476, + 0.005828829016536474, + 0.005935883615165949, + 0.0053807152435183525, + 0.01573878526687622, + 0.0035330653190612793, + 0.004266762174665928, + 0.003512964118272066, + 0.017395848408341408, + 0.008404775522649288, + 0.0036158435977995396, + 0.00530129624530673, + 0.022442985326051712, + 0.012426252476871014, + 0.01287509873509407, + 0.021104730665683746, + 0.01396999228745699, + 0.02229941263794899, + 0.012825840152800083, + 0.01330721564590931, + 0.006972267758101225, + 0.0069067408330738544, + 0.0038939486257731915, + 0.006684856489300728, + 0.005213518626987934, + 0.007359301205724478, + 0.007102211471647024, + 0.005280424375087023, + 0.03407344967126846, + 0.020483335480093956, + 0.017466513440012932, + 0.018019411712884903, + 0.014166761189699173, + 0.030760383233428, + 0.013160296715795994, + 0.011016509495675564, + 0.008803381584584713, + 0.009996904991567135, + 0.009083200246095657, + 0.012253333814442158, + 0.013613135553896427, + 0.01157283317297697, + 0.009905899874866009, + 0.007417886983603239, + 0.009339654818177223, + 0.006468718405812979, + 0.005095832049846649, + 0.005697925575077534, + 0.005604919977486134, + 0.010847847908735275, + 0.005963491275906563, + 0.007998992688953876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.025766756385564804 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260316655039787 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260316655039787, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260316655039787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19380255043506622, + 0.06274700909852982, + 0.03948283940553665, + 0.053851086646318436, + 0.039349813014268875, + 0.12815390527248383, + 0.09402793645858765, + 0.16988593339920044, + 0.18069125711917877, + 0.03800756484270096 + ], + "credit_map": [ + 0.017400380223989487, + 0.06122252717614174, + 0.020318759605288506, + 0.020895417779684067, + 0.023321012035012245, + 0.015552214346826077, + 0.019866731017827988, + 0.015225524082779884, + 0.005815383046865463, + 0.01118324976414442, + 0.006231042556464672, + 0.007488344330340624, + 0.012086184695363045, + 0.008732304908335209, + 0.005324712023139, + 0.005885791499167681, + 0.003895699745044112, + 0.006476457230746746, + 0.004505007993429899, + 0.010238513350486755, + 0.0030394424684345722, + 0.00300230854190886, + 0.004136745352298021, + 0.004188669845461845, + 0.00669475132599473, + 0.00870100874453783, + 0.0059816003777086735, + 0.0068466514348983765, + 0.007458025589585304, + 0.003773055737838149, + 0.005368383601307869, + 0.009027618914842606, + 0.005070314276963472, + 0.0073386551812291145, + 0.003728819778189063, + 0.0043930113315582275, + 0.006652783136814833, + 0.0035912131424993277, + 0.0039367517456412315, + 0.004638270474970341, + 0.023636234924197197, + 0.012423264794051647, + 0.008489623665809631, + 0.014589791186153889, + 0.017298808321356773, + 0.023606907576322556, + 0.01874244585633278, + 0.009366839192807674, + 0.013485845178365707, + 0.0216832235455513, + 0.007732767146080732, + 0.012259403243660927, + 0.012019569985568523, + 0.010049596428871155, + 0.009335526265203953, + 0.007462017238140106, + 0.03606170415878296, + 0.02468212880194187, + 0.01344955526292324, + 0.0184298288077116, + 0.015913475304841995, + 0.03378360718488693, + 0.013638868927955627, + 0.013926785439252853, + 0.020376022905111313, + 0.02011132799088955, + 0.023264482617378235, + 0.014972787350416183, + 0.03725232928991318, + 0.0233094971626997, + 0.0209407489746809, + 0.020464083179831505, + 0.0050553372129797935, + 0.004870337434113026, + 0.004488746169954538, + 0.004500347655266523, + 0.004499431233853102, + 0.004896258469671011, + 0.004284801427274942, + 0.0054123057052493095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005260316655039787 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786837734282017 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786837734282017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786837734282017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16377969086170197, + 0.043778933584690094, + 0.10917852073907852, + 0.09537413716316223, + 0.07973571121692657, + 0.11174388974905014, + 0.10708968341350555, + 0.11450134217739105, + 0.11300327628850937, + 0.06181488186120987 + ], + "credit_map": [ + 0.020121930167078972, + 0.03040481172502041, + 0.024759894236922264, + 0.022485876455903053, + 0.01816066913306713, + 0.016317371279001236, + 0.019195102155208588, + 0.012334039434790611, + 0.004328721668571234, + 0.009197384119033813, + 0.0057518561370670795, + 0.007758972700685263, + 0.004046741407364607, + 0.004327621776610613, + 0.003657958237454295, + 0.004709677770733833, + 0.015983212739229202, + 0.011182988993823528, + 0.011743299663066864, + 0.009074043482542038, + 0.01582654006779194, + 0.01517910324037075, + 0.014507564716041088, + 0.015681764110922813, + 0.008542041294276714, + 0.01720493659377098, + 0.013820095919072628, + 0.015609425492584705, + 0.016302699223160744, + 0.007000462152063847, + 0.00737611623480916, + 0.00951835885643959, + 0.02137972228229046, + 0.006294744089245796, + 0.005163447465747595, + 0.0048921192064881325, + 0.028211673721671104, + 0.004870388191193342, + 0.0048380144871771336, + 0.004085599910467863, + 0.01893225871026516, + 0.01036078017205, + 0.013064764440059662, + 0.014411122538149357, + 0.011904656887054443, + 0.018551472574472427, + 0.01167951337993145, + 0.012839319184422493, + 0.014724786393344402, + 0.028204068541526794, + 0.008409246802330017, + 0.01585233211517334, + 0.012458986602723598, + 0.011027844622731209, + 0.010309476405382156, + 0.0061029368080198765, + 0.019999470561742783, + 0.011887001805007458, + 0.010237331502139568, + 0.013591739349067211, + 0.012786825187504292, + 0.02316461317241192, + 0.011326042003929615, + 0.011508315801620483, + 0.015159270726144314, + 0.010502231307327747, + 0.015606023371219635, + 0.009948905557394028, + 0.022572945803403854, + 0.014796347357332706, + 0.012506764382123947, + 0.011910785920917988, + 0.01094998512417078, + 0.004659338388592005, + 0.008154147304594517, + 0.004434437956660986, + 0.004896495025604963, + 0.015583863481879234, + 0.006179612595587969, + 0.00695700291544199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009786837734282017 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028181858360767365 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.0028181858360767365, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028181858360767365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0805012509226799, + 0.04125582054257393, + 0.10374266654253006, + 0.034009382128715515, + 0.052823569625616074, + 0.1418791115283966, + 0.07070405781269073, + 0.14083366096019745, + 0.2583426833152771, + 0.07590781897306442 + ], + "credit_map": [ + 0.014169457368552685, + 0.011283536441624165, + 0.006623308174312115, + 0.007457173895090818, + 0.0077415937557816505, + 0.012017512694001198, + 0.009852690622210503, + 0.01135597936809063, + 0.004933165851980448, + 0.006536463741213083, + 0.005525462329387665, + 0.005087793804705143, + 0.004188908729702234, + 0.004418190103024244, + 0.0048398361541330814, + 0.005726000294089317, + 0.014696027152240276, + 0.010504976846277714, + 0.010546397417783737, + 0.012431778013706207, + 0.017492635175585747, + 0.01363818533718586, + 0.011689110659062862, + 0.012743554078042507, + 0.004006220959126949, + 0.0033544329926371574, + 0.004195660352706909, + 0.003965305630117655, + 0.004551393445581198, + 0.003017532406374812, + 0.004132230766117573, + 0.00678660674020648, + 0.007778744678944349, + 0.004429957829415798, + 0.006472595501691103, + 0.003836272284388542, + 0.010168480686843395, + 0.004757644608616829, + 0.005346226971596479, + 0.010033650323748589, + 0.020305093377828598, + 0.017314504832029343, + 0.009091691114008427, + 0.014166200533509254, + 0.019435081630945206, + 0.0175539031624794, + 0.01742580533027649, + 0.026586828753352165, + 0.01188271027058363, + 0.008544464595615864, + 0.007456379942595959, + 0.0075140222907066345, + 0.00653338897973299, + 0.011371313594281673, + 0.007368486374616623, + 0.01003329548984766, + 0.028255106881260872, + 0.013881241902709007, + 0.01636960729956627, + 0.02038375847041607, + 0.01640726439654827, + 0.02156318910419941, + 0.012402893975377083, + 0.011570601724088192, + 0.028876347467303276, + 0.027650589123368263, + 0.03095969371497631, + 0.02427808567881584, + 0.04967310279607773, + 0.03523550555109978, + 0.032275233417749405, + 0.02939411625266075, + 0.012516241520643234, + 0.008564168587327003, + 0.00844335462898016, + 0.009907642379403114, + 0.007401794195175171, + 0.009803535416722298, + 0.008661340922117233, + 0.010609735734760761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.0028181858360767365 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058513894677162 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.016058513894677162, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058513894677162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12707698345184326, + 0.06120960786938667, + 0.06918993592262268, + 0.09825403243303299, + 0.11817452311515808, + 0.09298934787511826, + 0.07457531243562698, + 0.08394509553909302, + 0.16990602016448975, + 0.1046791598200798 + ], + "credit_map": [ + 0.021933021023869514, + 0.015356650575995445, + 0.01030477974563837, + 0.01229042187333107, + 0.010035947896540165, + 0.015553970821201801, + 0.01907947286963463, + 0.022522715851664543, + 0.0059714228846132755, + 0.011904807761311531, + 0.0061122458428144455, + 0.009237469173967838, + 0.010158820077776909, + 0.006192074157297611, + 0.005379913374781609, + 0.006252853665500879, + 0.008650587871670723, + 0.009631196968257427, + 0.006877008359879255, + 0.008728920482099056, + 0.014493669383227825, + 0.00836099311709404, + 0.00584713090211153, + 0.006600424647331238, + 0.01256874855607748, + 0.008548070676624775, + 0.014483697712421417, + 0.013529720716178417, + 0.012016816064715385, + 0.011108235456049442, + 0.01123660709708929, + 0.014762134291231632, + 0.03548383712768555, + 0.009114937856793404, + 0.007702939677983522, + 0.005203684791922569, + 0.04398272931575775, + 0.006862987764179707, + 0.005761727225035429, + 0.0040616681799292564, + 0.01499086245894432, + 0.009280648082494736, + 0.010910226963460445, + 0.010569044388830662, + 0.009277578443288803, + 0.01801113225519657, + 0.009823482483625412, + 0.010126370936632156, + 0.011098824441432953, + 0.016370410099625587, + 0.007460242137312889, + 0.01003581378608942, + 0.008196322247385979, + 0.006845044903457165, + 0.007841887883841991, + 0.006726761814206839, + 0.015922535210847855, + 0.012886502780020237, + 0.006273928098380566, + 0.010579737834632397, + 0.008447611704468727, + 0.01408067811280489, + 0.010443788021802902, + 0.0053103151731193066, + 0.021930471062660217, + 0.02043178491294384, + 0.027203159406781197, + 0.015021948143839836, + 0.028182461857795715, + 0.02418912574648857, + 0.01734582707285881, + 0.015601243823766708, + 0.01533602550625801, + 0.013907034881412983, + 0.006929174531251192, + 0.012718242593109608, + 0.012808863073587418, + 0.015974711626768112, + 0.011372682638466358, + 0.015632420778274536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.016058513894677162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, rendered in a vibrant animation style with bold silhouettes and highly saturated colors, follows a classic American cruiser trike around a 3,000-meter steeplechase course. The trike features a gleaming cherry-red frame, wide, swept-back chrome handlebars, and a large, padded black saddle. Iridescent satin streamers tied to the handlebar grips and rear axle flutter wildly in its wake. The scene is set against a brilliant blue sky, where a large, stylized cartoon sun with a smiling face radiates a warm, cheerful glow. As the trike approaches a water obstacle, the camera remains low, emphasizing the impending impact. The trike dramatically splashes into the vibrant, azure water of the Danger Pit, its large front wheel impacting the surface and launching a perfect, shimmering arc of water into the air. Each cartoon-like droplet in the spray catches the sunlight, shimmering with magical intensity before falling back down, creating a moment of slow-motion, joyful wonder amidst the dynamic race.", + "seed": 27926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_7413_seed_27926_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": -0.008608166640624405, + "vlm_logprob": -0.008608166640624405, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.008608166640624405, + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013350967317819595 + ], + "opposite_logprob_tokens": [ + -4.388350963592529 + ], + "logprob_reward": -0.013350967317819595, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013350967317819595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05129716917872429, + 0.12363220006227493, + 0.11361720412969589, + 0.4233335256576538, + 0.03525104001164436, + 0.07993081212043762, + 0.05367834493517876, + 0.0357554629445076, + 0.041010089218616486, + 0.04249407723546028 + ], + "credit_map": [ + 0.007277850527316332, + 0.006837309338152409, + 0.0062264855951070786, + 0.005280615296214819, + 0.006581210996955633, + 0.006380369886755943, + 0.0070402477867901325, + 0.005673081614077091, + 0.019650757312774658, + 0.009646883234381676, + 0.007187975104898214, + 0.005191884469240904, + 0.013427192345261574, + 0.026571352034807205, + 0.010656759142875671, + 0.0312994010746479, + 0.01617530919611454, + 0.009460615925490856, + 0.017316438257694244, + 0.010154854506254196, + 0.008894925005733967, + 0.013674287125468254, + 0.011422445066273212, + 0.0265183225274086, + 0.10458105802536011, + 0.024594726040959358, + 0.021818559616804123, + 0.018548043444752693, + 0.0269476231187582, + 0.11310091614723206, + 0.042054433375597, + 0.0716881975531578, + 0.0038660671561956406, + 0.0036995019763708115, + 0.0026749649550765753, + 0.0031431333627551794, + 0.0054196761921048164, + 0.003846034873276949, + 0.0054093776270747185, + 0.007192282471805811, + 0.0075655668042600155, + 0.007637629751116037, + 0.010314513929188251, + 0.008881313726305962, + 0.011818083934485912, + 0.009324158541858196, + 0.010886934585869312, + 0.013502610847353935, + 0.012739678844809532, + 0.01817513071000576, + 0.004001074004918337, + 0.0026783738285303116, + 0.0036938521079719067, + 0.00597681337967515, + 0.003347644815221429, + 0.0030657805036753416, + 0.006401068065315485, + 0.004271905403584242, + 0.004522455390542746, + 0.005241753999143839, + 0.0038514442276209593, + 0.004348150920122862, + 0.0034805317409336567, + 0.003638150403276086, + 0.007208623923361301, + 0.0066028255969285965, + 0.005543064326047897, + 0.0066587794572114944, + 0.0044051483273506165, + 0.002957942895591259, + 0.003755694255232811, + 0.0038780111353844404, + 0.0052349865436553955, + 0.006781389936804771, + 0.006590340752154589, + 0.005507260095328093, + 0.006849291734397411, + 0.0027151070535182953, + 0.005571629852056503, + 0.0032440691720694304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.013350967317819595 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04309411719441414, + 0.08858897536993027, + 0.09431172162294388, + 0.38872215151786804, + 0.014403965324163437, + 0.033423732966184616, + 0.1200936883687973, + 0.10800886154174805, + 0.046749893575906754, + 0.06260285526514053 + ], + "credit_map": [ + 0.005768457427620888, + 0.00550855090841651, + 0.005546477623283863, + 0.0050308313220739365, + 0.0053720055148005486, + 0.005523022264242172, + 0.004519536625593901, + 0.005825236439704895, + 0.013861311599612236, + 0.007362560369074345, + 0.005874749273061752, + 0.0032318360172212124, + 0.009908409789204597, + 0.018783608451485634, + 0.007562498562037945, + 0.022004006430506706, + 0.013216360472142696, + 0.01043460238724947, + 0.01521132979542017, + 0.009441048838198185, + 0.006290757097303867, + 0.00987691804766655, + 0.0092540942132473, + 0.020586613565683365, + 0.1426963061094284, + 0.015897877514362335, + 0.010223405435681343, + 0.010060268454253674, + 0.01347892265766859, + 0.16674301028251648, + 0.015113144181668758, + 0.014509213156998158, + 0.0018248881679028273, + 0.0017041759565472603, + 0.0011879887897521257, + 0.0013780688168480992, + 0.0024312029127031565, + 0.0014823260717093945, + 0.00242465827614069, + 0.0019706578459590673, + 0.004654580727219582, + 0.003080185502767563, + 0.002935553900897503, + 0.0033133206889033318, + 0.005762847140431404, + 0.0033166073262691498, + 0.005404154770076275, + 0.004956486169248819, + 0.03378050774335861, + 0.048000041395425797, + 0.01128931250423193, + 0.004751833621412516, + 0.006521214731037617, + 0.006756150629371405, + 0.004571129102259874, + 0.004423505160957575, + 0.015224674716591835, + 0.011923493817448616, + 0.010351843200623989, + 0.014774139970541, + 0.014341229572892189, + 0.0187500212341547, + 0.010300815105438232, + 0.01234265137463808, + 0.007735513616353273, + 0.0070478604175150394, + 0.006623282562941313, + 0.00808445643633604, + 0.004988330882042646, + 0.004426002502441406, + 0.003949072677642107, + 0.0038953775074332952, + 0.008367536589503288, + 0.009039226919412613, + 0.010433021001517773, + 0.012149649672210217, + 0.008428522385656834, + 0.0035663763992488384, + 0.006067253183573484, + 0.004551274236291647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.002491467399522662 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11118525266647339, + 0.07868916541337967, + 0.1482958197593689, + 0.18255193531513214, + 0.05271928012371063, + 0.066273532807827, + 0.07138339430093765, + 0.07775261998176575, + 0.1107669249176979, + 0.10038211196660995 + ], + "credit_map": [ + 0.01780279539525509, + 0.02868511527776718, + 0.01271167304366827, + 0.011543918401002884, + 0.013407078571617603, + 0.009895686991512775, + 0.008751326240599155, + 0.008387651294469833, + 0.01117786206305027, + 0.011023628525435925, + 0.0055008940398693085, + 0.003891122294589877, + 0.011186541989445686, + 0.01390805933624506, + 0.008077493868768215, + 0.01392356213182211, + 0.02086426131427288, + 0.010731860995292664, + 0.025968607515096664, + 0.01661810651421547, + 0.010840017348527908, + 0.0157318115234375, + 0.013836721889674664, + 0.03370443731546402, + 0.062319930642843246, + 0.008126305416226387, + 0.006492448505014181, + 0.007773431483656168, + 0.006546365097165108, + 0.07296392321586609, + 0.010106738656759262, + 0.008222782053053379, + 0.007881547324359417, + 0.007386552169919014, + 0.0034008086659014225, + 0.005324073601514101, + 0.010198213160037994, + 0.005634302273392677, + 0.005417207721620798, + 0.007476574741303921, + 0.015850022435188293, + 0.00620771711692214, + 0.006412866525352001, + 0.006794461514800787, + 0.009015548042953014, + 0.0058284965343773365, + 0.008246297016739845, + 0.007918130606412888, + 0.013365981169044971, + 0.016942109912633896, + 0.014613792300224304, + 0.0053169261664152145, + 0.0033370014280080795, + 0.008341803215444088, + 0.004561630543321371, + 0.004904148634523153, + 0.010169146582484245, + 0.008739294484257698, + 0.008107424713671207, + 0.010117332451045513, + 0.009967333637177944, + 0.013509558513760567, + 0.008239535614848137, + 0.00890299491584301, + 0.01659442111849785, + 0.016176003962755203, + 0.014676790684461594, + 0.02106037549674511, + 0.011120639741420746, + 0.011164885014295578, + 0.010596316307783127, + 0.00937749445438385, + 0.012807748280465603, + 0.014062420465052128, + 0.017023855820298195, + 0.020933715626597404, + 0.014442765153944492, + 0.004683731589466333, + 0.010079423896968365, + 0.0063484590500593185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.014193654991686344 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019386084750294685 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019386084750294685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019386084750294685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0730101466178894, + 0.07163062691688538, + 0.13854321837425232, + 0.06923051178455353, + 0.017779206857085228, + 0.042233459651470184, + 0.14475314319133759, + 0.11160315573215485, + 0.18848678469657898, + 0.14272969961166382 + ], + "credit_map": [ + 0.009696749970316887, + 0.01661710813641548, + 0.010376650840044022, + 0.007688866928219795, + 0.009224629029631615, + 0.0062693338841199875, + 0.007372673135250807, + 0.005764134228229523, + 0.01158592663705349, + 0.008569285273551941, + 0.007339486852288246, + 0.004714540205895901, + 0.0038313085678964853, + 0.01018141396343708, + 0.005183251574635506, + 0.020225418731570244, + 0.021191883832216263, + 0.008474242873489857, + 0.01647588610649109, + 0.011853847652673721, + 0.009485448710620403, + 0.01902674324810505, + 0.010710309259593487, + 0.041324857622385025, + 0.012087364681065083, + 0.0051403651013970375, + 0.004305921960622072, + 0.004631008952856064, + 0.005856846459209919, + 0.011274144053459167, + 0.010019371286034584, + 0.015915486961603165, + 0.002319708000868559, + 0.0025070232804864645, + 0.0011246439535170794, + 0.001292258850298822, + 0.002527548698708415, + 0.0020226244814693928, + 0.003206247929483652, + 0.002779152011498809, + 0.00664774002507329, + 0.004444411024451256, + 0.004056698642671108, + 0.003966646734625101, + 0.00590684125199914, + 0.003973805345594883, + 0.006908005103468895, + 0.00632931012660265, + 0.040908802300691605, + 0.05529182404279709, + 0.02075791358947754, + 0.004470163490623236, + 0.0076402765698730946, + 0.006507229991257191, + 0.0045493715442717075, + 0.004627563990652561, + 0.019307250156998634, + 0.014983098022639751, + 0.016858363524079323, + 0.01880243420600891, + 0.01100349985063076, + 0.011393814347684383, + 0.010707619599997997, + 0.008547080680727959, + 0.02930164523422718, + 0.028947738930583, + 0.024202438071370125, + 0.03072911687195301, + 0.022869886830449104, + 0.019758202135562897, + 0.0169468205422163, + 0.015730945393443108, + 0.014934362843632698, + 0.018065718933939934, + 0.026582637801766396, + 0.031518012285232544, + 0.017881743609905243, + 0.009197861887514591, + 0.01363238226622343, + 0.01091697160154581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0019386084750294685 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106613501906395 + ], + "opposite_logprob_tokens": [ + -4.51106595993042 + ], + "logprob_reward": -0.01106613501906395, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106613501906395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07214165478944778, + 0.1391974538564682, + 0.1706814020872116, + 0.14332899451255798, + 0.027665477246046066, + 0.03386364132165909, + 0.05912807583808899, + 0.08842527866363525, + 0.18604126572608948, + 0.07952671498060226 + ], + "credit_map": [ + 0.011693124659359455, + 0.011715489439666271, + 0.009446087293326855, + 0.006730124354362488, + 0.009474048390984535, + 0.00842156819999218, + 0.008034203201532364, + 0.0066270167008042336, + 0.01974949613213539, + 0.013886695727705956, + 0.01454074028879404, + 0.010515117086470127, + 0.015445953235030174, + 0.022134358063340187, + 0.016372617334127426, + 0.026552483439445496, + 0.02300925925374031, + 0.01763487607240677, + 0.018764261156320572, + 0.02024909481406212, + 0.017798524349927902, + 0.019183527678251266, + 0.016236985102295876, + 0.03780487924814224, + 0.04011505842208862, + 0.010516676120460033, + 0.005963529925793409, + 0.005984497722238302, + 0.007882063277065754, + 0.04438534379005432, + 0.0067792972549796104, + 0.02170254848897457, + 0.003963332157582045, + 0.004300021566450596, + 0.0016229323809966445, + 0.0018277009949088097, + 0.005393815692514181, + 0.0031261721160262823, + 0.0039898911491036415, + 0.0034416140988469124, + 0.008088010363280773, + 0.0030855785589665174, + 0.0035394628066569567, + 0.0026729812379926443, + 0.003830669680610299, + 0.004523773677647114, + 0.004621160216629505, + 0.0035020066425204277, + 0.010135533288121223, + 0.005057100672274828, + 0.020311320200562477, + 0.0035195061936974525, + 0.0031776600517332554, + 0.007252740673720837, + 0.0036770121660083532, + 0.005997209344059229, + 0.012524244375526905, + 0.011664969846606255, + 0.015138093382120132, + 0.013241060078144073, + 0.006780738476663828, + 0.012991749681532383, + 0.008610161021351814, + 0.007474263198673725, + 0.02355937659740448, + 0.02755642682313919, + 0.018207518383860588, + 0.02126486413180828, + 0.027951812371611595, + 0.03311654180288315, + 0.0179771538823843, + 0.01640758477151394, + 0.010047556832432747, + 0.012333248741924763, + 0.012203577905893326, + 0.009055908769369125, + 0.012510173954069614, + 0.007126801647245884, + 0.009541076608002186, + 0.006708374246954918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.01106613501906395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, rendered in a vibrant animation style with bold silhouettes and highly saturated colors, follows a classic American cruiser trike around a 3,000-meter steeplechase course. The trike features a gleaming cherry-red frame, wide, swept-back chrome handlebars, and a large, padded black saddle. Iridescent satin streamers tied to the handlebar grips and rear axle flutter wildly in its wake. The scene is set against a brilliant blue sky, where a large, stylized cartoon sun with a smiling face radiates a warm, cheerful glow. As the trike approaches a water obstacle, the camera remains low, emphasizing the impending impact. The trike dramatically splashes into the vibrant, azure water of the Danger Pit, its large front wheel impacting the surface and launching a perfect, shimmering arc of water into the air. Each cartoon-like droplet in the spray catches the sunlight, shimmering with magical intensity before falling back down, creating a moment of slow-motion, joyful wonder amidst the dynamic race.", + "seed": 27927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_7413_seed_27927_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": -0.006709455698728562, + "vlm_logprob": -0.006709455698728562, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.006709455698728562, + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047088684514164925 + ], + "opposite_logprob_tokens": [ + -5.754708766937256 + ], + "logprob_reward": -0.0047088684514164925, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0047088684514164925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660691440105438, + 0.07403258979320526, + 0.024723106995224953, + 0.03757631778717041, + 0.05907512456178665, + 0.04508132487535477, + 0.06690668314695358, + 0.24978679418563843, + 0.0693713054060936, + 0.10737767815589905 + ], + "credit_map": [ + 0.02109498716890812, + 0.02934025414288044, + 0.027726823464035988, + 0.025416608899831772, + 0.0639999657869339, + 0.031826842576265335, + 0.044455986469984055, + 0.022207673639059067, + 0.006705532781779766, + 0.0071858251467347145, + 0.009103798307478428, + 0.009507658891379833, + 0.013802657835185528, + 0.009986097924411297, + 0.00901949591934681, + 0.008721521124243736, + 0.0031500256154686213, + 0.003244521329179406, + 0.0034010405652225018, + 0.002978474833071232, + 0.002926358487457037, + 0.0033668505493551493, + 0.002662458224222064, + 0.0029933787882328033, + 0.006837757304310799, + 0.004771160893142223, + 0.004653791896998882, + 0.004173700697720051, + 0.0037512837443500757, + 0.002589333103969693, + 0.007231800816953182, + 0.0035674890968948603, + 0.011608028784394264, + 0.008086319081485271, + 0.005680636968463659, + 0.005544441286474466, + 0.007161616813391447, + 0.0054961866699159145, + 0.00824833009392023, + 0.007249563466757536, + 0.005204210989177227, + 0.003919994458556175, + 0.0053457519970834255, + 0.006723479833453894, + 0.005389944650232792, + 0.0046809157356619835, + 0.008369640447199345, + 0.0054473876953125, + 0.009028645232319832, + 0.006200271658599377, + 0.005358874332159758, + 0.007314469199627638, + 0.006741114892065525, + 0.013063053600490093, + 0.006469381973147392, + 0.012730875983834267, + 0.04547988995909691, + 0.031334441155195236, + 0.05213882029056549, + 0.02434643730521202, + 0.018766334280371666, + 0.025423718616366386, + 0.022717760875821114, + 0.029579371213912964, + 0.006678873673081398, + 0.005034234374761581, + 0.011042394675314426, + 0.00967342033982277, + 0.0072250207886099815, + 0.01472477987408638, + 0.008514202199876308, + 0.006478378549218178, + 0.013531490229070187, + 0.01130917388945818, + 0.016219474375247955, + 0.01682380400598049, + 0.009655189700424671, + 0.006760660093277693, + 0.019548600539565086, + 0.013529288582503796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0047088684514164925 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2502809464931488, + 0.06398314982652664, + 0.0672871321439743, + 0.04003685340285301, + 0.1053091511130333, + 0.027829518541693687, + 0.0306067056953907, + 0.144206240773201, + 0.06222698092460632, + 0.20823335647583008 + ], + "credit_map": [ + 0.020696518942713737, + 0.03626544401049614, + 0.022670764476060867, + 0.026308339089155197, + 0.06320743262767792, + 0.025701064616441727, + 0.03716439753770828, + 0.018266959115862846, + 0.006287878844887018, + 0.008074427023530006, + 0.006718805059790611, + 0.007371306419372559, + 0.009480331093072891, + 0.008686171844601631, + 0.007707263808697462, + 0.00965695921331644, + 0.01240494754165411, + 0.005820798221975565, + 0.009035514667630196, + 0.007749658077955246, + 0.007776352576911449, + 0.010603666305541992, + 0.005503600463271141, + 0.008392591960728168, + 0.006788409780710936, + 0.004665826912969351, + 0.007617729250341654, + 0.004284061025828123, + 0.0037413383834064007, + 0.002871780889108777, + 0.006810768041759729, + 0.0032569398172199726, + 0.025953322649002075, + 0.012243813835084438, + 0.01589476875960827, + 0.010111109353601933, + 0.009384200908243656, + 0.011028217151761055, + 0.011690784245729446, + 0.009002936072647572, + 0.003869006410241127, + 0.004036580678075552, + 0.0034021339379251003, + 0.0038494498003274202, + 0.002933814190328121, + 0.0025470207910984755, + 0.002804865362122655, + 0.004386644810438156, + 0.00485872570425272, + 0.002818305278196931, + 0.0031707189045846462, + 0.0036566341295838356, + 0.002759164897724986, + 0.00471840426325798, + 0.0038026338443160057, + 0.004822117276489735, + 0.01454372238367796, + 0.020167626440525055, + 0.030434997752308846, + 0.017792727798223495, + 0.014401175081729889, + 0.016641991212964058, + 0.017714673653244972, + 0.012509303167462349, + 0.006176888011395931, + 0.005496455822139978, + 0.009693127125501633, + 0.011192020028829575, + 0.007256006821990013, + 0.008834995329380035, + 0.008591191843152046, + 0.004986290354281664, + 0.02589564397931099, + 0.02284749038517475, + 0.023965885862708092, + 0.024886267259716988, + 0.017322149127721786, + 0.017713887616991997, + 0.04543038457632065, + 0.03017163649201393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.004098350182175636 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1819155216217041, + 0.09848164767026901, + 0.11983026564121246, + 0.049383874982595444, + 0.04717417061328888, + 0.05828935280442238, + 0.07280119508504868, + 0.20118927955627441, + 0.07934913039207458, + 0.09158556908369064 + ], + "credit_map": [ + 0.016854355111718178, + 0.027617717161774635, + 0.017415111884474754, + 0.020058372989296913, + 0.036911070346832275, + 0.019542070105671883, + 0.023983627557754517, + 0.01953318901360035, + 0.010476257652044296, + 0.010134199634194374, + 0.010580433532595634, + 0.010020794346928596, + 0.01938382349908352, + 0.013711294159293175, + 0.01366501022130251, + 0.010509833693504333, + 0.02270430512726307, + 0.013054490089416504, + 0.01900804042816162, + 0.01038585975766182, + 0.014828274957835674, + 0.019983312115073204, + 0.0075200945138931274, + 0.012345886789262295, + 0.00712982565164566, + 0.004898963030427694, + 0.010044421069324017, + 0.005254268180578947, + 0.003662723582237959, + 0.0035481734666973352, + 0.010121402330696583, + 0.004724099300801754, + 0.008307039737701416, + 0.006355874706059694, + 0.0034943276550620794, + 0.004433685913681984, + 0.005051604472100735, + 0.006843807175755501, + 0.006613706238567829, + 0.006074125878512859, + 0.006213061977177858, + 0.006302612368017435, + 0.006842262577265501, + 0.007422742433845997, + 0.006407252978533506, + 0.006689010187983513, + 0.010017436929047108, + 0.008394970558583736, + 0.007103694602847099, + 0.0043527791276574135, + 0.004925793036818504, + 0.007004174403846264, + 0.0047113364562392235, + 0.018727079033851624, + 0.0056558633223176, + 0.020320475101470947, + 0.02552505023777485, + 0.022988364100456238, + 0.046850670129060745, + 0.021465329453349113, + 0.018221810460090637, + 0.02513747103512287, + 0.02237301506102085, + 0.018627557903528214, + 0.009579786099493504, + 0.008045253343880177, + 0.012676404789090157, + 0.013437176123261452, + 0.008220234885811806, + 0.011571325361728668, + 0.010468030348420143, + 0.005350918509066105, + 0.013128419406712055, + 0.009961468167603016, + 0.012850206345319748, + 0.011004510335624218, + 0.006528940983116627, + 0.0066335308365523815, + 0.019164256751537323, + 0.012314235791563988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.016071418300271034 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019388464279472828 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019388464279472828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019388464279472828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13929973542690277, + 0.032662227749824524, + 0.02052018605172634, + 0.03593740612268448, + 0.029635164886713028, + 0.04522084817290306, + 0.052438851445913315, + 0.2053021937608719, + 0.061522696167230606, + 0.3774607181549072 + ], + "credit_map": [ + 0.01485470961779356, + 0.01411929726600647, + 0.012059055268764496, + 0.010770801454782486, + 0.0359441377222538, + 0.018054785206913948, + 0.01943117193877697, + 0.014065776020288467, + 0.003375428030267358, + 0.003996615298092365, + 0.004117525648325682, + 0.002845832845196128, + 0.004902799613773823, + 0.005337257403880358, + 0.0036113467067480087, + 0.004475423600524664, + 0.002667526016011834, + 0.002198379021137953, + 0.0027312468737363815, + 0.00235807616263628, + 0.0031373659148812294, + 0.0032167837489396334, + 0.0019816060084849596, + 0.0022292037028819323, + 0.008175327442586422, + 0.0038115845527499914, + 0.00499261450022459, + 0.0031562065705657005, + 0.0033229703549295664, + 0.0024505204055458307, + 0.006694141775369644, + 0.003334041452035308, + 0.007649899460375309, + 0.003203080967068672, + 0.004561726935207844, + 0.0023776565212756395, + 0.003285985207185149, + 0.0027616824954748154, + 0.002870220923796296, + 0.0029249130748212337, + 0.006492981221526861, + 0.008188568986952305, + 0.0071984222158789635, + 0.004961303435266018, + 0.004600862972438335, + 0.0035705682821571827, + 0.0031830123625695705, + 0.0070251282304525375, + 0.006683955900371075, + 0.003949887119233608, + 0.004661117680370808, + 0.0051817260682582855, + 0.006311063654720783, + 0.0066435569897294044, + 0.006637439131736755, + 0.01237010769546032, + 0.026434512808918953, + 0.018386295065283775, + 0.05858709663152695, + 0.021011169999837875, + 0.012582682073116302, + 0.026236537843942642, + 0.019410284236073494, + 0.02265363559126854, + 0.0069496710784733295, + 0.00547022232785821, + 0.009498311206698418, + 0.010394816286861897, + 0.006438549142330885, + 0.009441851638257504, + 0.008859721012413502, + 0.00446955393999815, + 0.04045966640114784, + 0.03942887857556343, + 0.044533032923936844, + 0.04248128458857536, + 0.030903955921530724, + 0.04026422277092934, + 0.08469555526971817, + 0.054694097489118576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0019388464279472828 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006729795131832361 + ], + "opposite_logprob_tokens": [ + -5.006729602813721 + ], + "logprob_reward": -0.006729795131832361, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006729795131832361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10683111101388931, + 0.07469745725393295, + 0.04441789537668228, + 0.0493394210934639, + 0.1380099654197693, + 0.045074593275785446, + 0.05523376911878586, + 0.21916349232196808, + 0.04893180727958679, + 0.2183004766702652 + ], + "credit_map": [ + 0.012167781591415405, + 0.010526028461754322, + 0.008949246257543564, + 0.010779174044728279, + 0.023071104660630226, + 0.013162890449166298, + 0.014704879373311996, + 0.013470010831952095, + 0.009178999811410904, + 0.00675106979906559, + 0.009292921051383018, + 0.008642749860882759, + 0.014826515689492226, + 0.009902977384626865, + 0.006790074519813061, + 0.009312145411968231, + 0.00667213648557663, + 0.005418222863227129, + 0.006376414094120264, + 0.004874400794506073, + 0.0048139640130102634, + 0.00740563590079546, + 0.003706154180690646, + 0.005150969605892897, + 0.01117138285189867, + 0.006058784667402506, + 0.005214451812207699, + 0.004867569077759981, + 0.003933990839868784, + 0.004884023685008287, + 0.008074114099144936, + 0.005135104060173035, + 0.03657485544681549, + 0.0169232040643692, + 0.019275780767202377, + 0.010357515886425972, + 0.01449910644441843, + 0.01622576080262661, + 0.011443077586591244, + 0.01271066628396511, + 0.004854497499763966, + 0.005243606865406036, + 0.00616389000788331, + 0.006729642394930124, + 0.00439990870654583, + 0.004064009990543127, + 0.006605539005249739, + 0.0070135029964149, + 0.006311940029263496, + 0.005487832240760326, + 0.007251366041600704, + 0.0073310937732458115, + 0.006370979361236095, + 0.006357109174132347, + 0.008510141633450985, + 0.007613308262079954, + 0.031223207712173462, + 0.015934409573674202, + 0.05125965550541878, + 0.02227817475795746, + 0.01466520968824625, + 0.02889842912554741, + 0.030008872970938683, + 0.02489555813372135, + 0.00838849600404501, + 0.004414599388837814, + 0.0065546222031116486, + 0.005342718679457903, + 0.006690110545605421, + 0.0059664552100002766, + 0.006644227541983128, + 0.004930578172206879, + 0.024280784651637077, + 0.021500151604413986, + 0.02455897629261017, + 0.02376740612089634, + 0.01897769421339035, + 0.02406097948551178, + 0.048711229115724564, + 0.03244326263666153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.006729795131832361 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot, rendered in a vibrant animation style with bold silhouettes and highly saturated colors, follows a classic American cruiser trike around a 3,000-meter steeplechase course. The trike features a gleaming cherry-red frame, wide, swept-back chrome handlebars, and a large, padded black saddle. Iridescent satin streamers tied to the handlebar grips and rear axle flutter wildly in its wake. The scene is set against a brilliant blue sky, where a large, stylized cartoon sun with a smiling face radiates a warm, cheerful glow. As the trike approaches a water obstacle, the camera remains low, emphasizing the impending impact. The trike dramatically splashes into the vibrant, azure water of the Danger Pit, its large front wheel impacting the surface and launching a perfect, shimmering arc of water into the air. Each cartoon-like droplet in the spray catches the sunlight, shimmering with magical intensity before falling back down, creating a moment of slow-motion, joyful wonder amidst the dynamic race.", + "seed": 27928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_8/index_7413_seed_27928_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": -0.010042789671570063, + "vlm_logprob": -0.010042789671570063, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.010042789671570063, + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968485958874226 + ], + "opposite_logprob_tokens": [ + -5.2559685707092285 + ], + "logprob_reward": -0.005968485958874226, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005968485958874226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22218108177185059, + 0.13217702507972717, + 0.07886211574077606, + 0.09443140774965286, + 0.12661679089069366, + 0.055714201182127, + 0.06019938364624977, + 0.12000709772109985, + 0.06015045568346977, + 0.04966041073203087 + ], + "credit_map": [ + 0.021441468968987465, + 0.0281978789716959, + 0.027690324932336807, + 0.0353253073990345, + 0.02677052468061447, + 0.022568725049495697, + 0.034493304789066315, + 0.025693558156490326, + 0.016643047332763672, + 0.0242457315325737, + 0.024072060361504555, + 0.016656870022416115, + 0.014904226176440716, + 0.013931198045611382, + 0.01072479784488678, + 0.010999075137078762, + 0.010361405089497566, + 0.013766760937869549, + 0.007930729538202286, + 0.009026254527270794, + 0.016682138666510582, + 0.008775864727795124, + 0.0063479263335466385, + 0.005971033126115799, + 0.012881485745310783, + 0.012020730413496494, + 0.008698029443621635, + 0.01126539520919323, + 0.013589035719633102, + 0.008485703729093075, + 0.013162857852876186, + 0.014328166842460632, + 0.02188764326274395, + 0.04124277085065842, + 0.010125325992703438, + 0.014772746711969376, + 0.01085793599486351, + 0.007921513170003891, + 0.01031481847167015, + 0.009494036436080933, + 0.013359527103602886, + 0.008514361456036568, + 0.004308519419282675, + 0.0029950649477541447, + 0.009482203051447868, + 0.006408604793250561, + 0.00534885423257947, + 0.005297064781188965, + 0.013513267040252686, + 0.014140365645289421, + 0.005715301726013422, + 0.004732704255729914, + 0.006632592063397169, + 0.006089845672249794, + 0.004831328988075256, + 0.004543979652225971, + 0.02184140309691429, + 0.03716962784528732, + 0.0074270982295274734, + 0.006678516510874033, + 0.015255842357873917, + 0.01207274291664362, + 0.00944104790687561, + 0.010120815597474575, + 0.004915822297334671, + 0.007191680837422609, + 0.006553965155035257, + 0.00795819703489542, + 0.007350512780249119, + 0.009506688453257084, + 0.009703753516077995, + 0.006969834677875042, + 0.00544360326603055, + 0.005711851175874472, + 0.005058967508375645, + 0.005720184184610844, + 0.007460536900907755, + 0.005682264920324087, + 0.007786727976053953, + 0.006796274334192276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005968485958874226 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15261925756931305, + 0.0734085887670517, + 0.11988403648138046, + 0.14519676566123962, + 0.05748060345649719, + 0.07874301820993423, + 0.08199642598628998, + 0.18403995037078857, + 0.047787655144929886, + 0.058843690901994705 + ], + "credit_map": [ + 0.019943974912166595, + 0.023817729204893112, + 0.018740201368927956, + 0.02765461802482605, + 0.01646875962615013, + 0.016825897619128227, + 0.01661546155810356, + 0.012552612461149693, + 0.008793644607067108, + 0.013883843086659908, + 0.010126161389052868, + 0.009020338766276836, + 0.00832002516835928, + 0.009744396433234215, + 0.006328263785690069, + 0.007191916462033987, + 0.017288682982325554, + 0.024844450876116753, + 0.009303495287895203, + 0.013133523054420948, + 0.028455713763833046, + 0.010334114544093609, + 0.00927856843918562, + 0.007245487067848444, + 0.022376269102096558, + 0.034017786383628845, + 0.013647058978676796, + 0.021788479760289192, + 0.01810765080153942, + 0.011990716680884361, + 0.009948689490556717, + 0.01332010980695486, + 0.00866222009062767, + 0.01573626697063446, + 0.004662682767957449, + 0.007374715059995651, + 0.005621789954602718, + 0.0035949177108705044, + 0.00538125354796648, + 0.006446756888180971, + 0.020514018833637238, + 0.013275024481117725, + 0.006490218918770552, + 0.004782208241522312, + 0.011114519089460373, + 0.008866111747920513, + 0.006628482602536678, + 0.007072439882904291, + 0.017446255311369896, + 0.013109276071190834, + 0.008082435466349125, + 0.006799308583140373, + 0.007524177897721529, + 0.01001160591840744, + 0.009303256869316101, + 0.009720106609165668, + 0.03312154486775398, + 0.06316529214382172, + 0.009724593721330166, + 0.010809196159243584, + 0.026463748887181282, + 0.01868678443133831, + 0.009326760657131672, + 0.012742041610181332, + 0.007513764314353466, + 0.006338087841868401, + 0.00488079572096467, + 0.005568724125623703, + 0.00816268753260374, + 0.00611911341547966, + 0.005324491299688816, + 0.0038799894973635674, + 0.006023714784532785, + 0.007437223102897406, + 0.00550051499158144, + 0.005229007452726364, + 0.009150107391178608, + 0.008465506136417389, + 0.007456995081156492, + 0.009580629877746105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0024903970770537853 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537237256765366 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537237256765366, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537237256765366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.192648783326149, + 0.05611696466803551, + 0.20860739052295685, + 0.09856536984443665, + 0.04871208593249321, + 0.09075971692800522, + 0.07682216912508011, + 0.11778345704078674, + 0.04714671149849892, + 0.0628373920917511 + ], + "credit_map": [ + 0.025857845321297646, + 0.03417109325528145, + 0.01758226938545704, + 0.039205629378557205, + 0.01875247247517109, + 0.019122062250971794, + 0.020404689013957977, + 0.017552712932229042, + 0.006580722518265247, + 0.010657846927642822, + 0.006216917186975479, + 0.0052031222730875015, + 0.007936784997582436, + 0.008805391378700733, + 0.004658343270421028, + 0.006057836581021547, + 0.030126294121146202, + 0.04197193309664726, + 0.023154424503445625, + 0.019648166373372078, + 0.043271254748106, + 0.016299385577440262, + 0.02207782305777073, + 0.012058109045028687, + 0.010755340568721294, + 0.015510940924286842, + 0.011872903443872929, + 0.019222017377614975, + 0.00988207571208477, + 0.008029812946915627, + 0.006646639667451382, + 0.016645634546875954, + 0.006434777285903692, + 0.0053305779583752155, + 0.003769588191062212, + 0.006544950418174267, + 0.005737080238759518, + 0.005776626989245415, + 0.0055330777540802956, + 0.009585407562553883, + 0.027014242485165596, + 0.015879014506936073, + 0.00954714696854353, + 0.005823230370879173, + 0.008852535858750343, + 0.009387184865772724, + 0.0073734004981815815, + 0.006882959511131048, + 0.015899674966931343, + 0.012172491289675236, + 0.007570066023617983, + 0.005924757104367018, + 0.007292695343494415, + 0.010769101791083813, + 0.00913548655807972, + 0.008057895116508007, + 0.019864609465003014, + 0.03728631138801575, + 0.006315858103334904, + 0.007363634184002876, + 0.016718432307243347, + 0.013368266634643078, + 0.008063698187470436, + 0.008802639320492744, + 0.008484563790261745, + 0.0058947959914803505, + 0.004651392810046673, + 0.0044595105573534966, + 0.009911693632602692, + 0.006010748445987701, + 0.003920737188309431, + 0.0038132681511342525, + 0.0066832005977630615, + 0.006589753087610006, + 0.009183833375573158, + 0.006329018156975508, + 0.009015253745019436, + 0.006600521504878998, + 0.009455513209104538, + 0.008980294689536095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.012537237256765366 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944311618804932 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944311618804932, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944311618804932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367330402135849, + 0.12844811379909515, + 0.064352847635746, + 0.1097443625330925, + 0.05275880917906761, + 0.13543681800365448, + 0.1562097668647766, + 0.09373442828655243, + 0.03456122800707817, + 0.08802052587270737 + ], + "credit_map": [ + 0.011470561847090721, + 0.01675213873386383, + 0.01205785758793354, + 0.030354194343090057, + 0.014550959691405296, + 0.012021269649267197, + 0.021686742082238197, + 0.017839327454566956, + 0.014333345927298069, + 0.023963581770658493, + 0.01877078413963318, + 0.015476135537028313, + 0.014554502442479134, + 0.013584902510046959, + 0.014300617389380932, + 0.013464255258440971, + 0.016167694702744484, + 0.008278689347207546, + 0.006169594824314117, + 0.006088364869356155, + 0.008157476782798767, + 0.010929776355624199, + 0.0038789059035480022, + 0.004682359751313925, + 0.015247099101543427, + 0.022687340155243874, + 0.008812651969492435, + 0.012233216315507889, + 0.020997673273086548, + 0.010119371116161346, + 0.008660194464027882, + 0.01098682265728712, + 0.007092484273016453, + 0.010649432428181171, + 0.004350100643932819, + 0.006407231092453003, + 0.008254955522716045, + 0.006176854483783245, + 0.005014420952647924, + 0.004813335835933685, + 0.032494526356458664, + 0.0203079916536808, + 0.014132796786725521, + 0.011078743264079094, + 0.02039140649139881, + 0.013103220611810684, + 0.012665778398513794, + 0.01126235630363226, + 0.02261565811932087, + 0.03440264239907265, + 0.010052659548819065, + 0.008785972371697426, + 0.020269032567739487, + 0.020783033221960068, + 0.021380696445703506, + 0.017920080572366714, + 0.018874190747737885, + 0.02861771173775196, + 0.006095461081713438, + 0.006303457543253899, + 0.009114997461438179, + 0.011071153916418552, + 0.007332650013267994, + 0.006324813235551119, + 0.004604132380336523, + 0.003560450626537204, + 0.0036467844620347023, + 0.003868765663355589, + 0.00611929502338171, + 0.004884230904281139, + 0.0045639690943062305, + 0.0033136000856757164, + 0.006318528670817614, + 0.011899484321475029, + 0.00836558174341917, + 0.009769291616976261, + 0.013264931738376617, + 0.01206404808908701, + 0.01448106486350298, + 0.011857603676617146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.005944311618804932 + }, + { + "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 classic American cruiser trike with a cherry-red frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the trike feature wide, swept-back chrome handlebars and a large padded black saddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent satin streamers tied to the handlebar grips and rear axle fluttering in the wake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the trike splash into an azure water obstacle known as the Danger Pit, creating a shimmering arc of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low and tracking the trike, emphasizing the impending impact with the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023273516446352005 + ], + "opposite_logprob_tokens": [ + -3.773273468017578 + ], + "logprob_reward": -0.023273516446352005, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023273516446352005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19275742769241333, + 0.07173297554254532, + 0.06095660850405693, + 0.1927918642759323, + 0.1218419075012207, + 0.08582796156406403, + 0.09013691544532776, + 0.10831840336322784, + 0.030243143439292908, + 0.045392733067274094 + ], + "credit_map": [ + 0.02275921404361725, + 0.015985211357474327, + 0.03136466071009636, + 0.021300489082932472, + 0.029459185898303986, + 0.025516165420413017, + 0.020982986316084862, + 0.025389540940523148, + 0.009238776750862598, + 0.010059479624032974, + 0.008044318296015263, + 0.010491898283362389, + 0.007247092667967081, + 0.00884968787431717, + 0.009223280474543571, + 0.008578444831073284, + 0.0075857434421777725, + 0.01082459744066, + 0.005231523886322975, + 0.006350313313305378, + 0.012934774160385132, + 0.006126485299319029, + 0.005395672749727964, + 0.006507506128400564, + 0.02560865506529808, + 0.03550748527050018, + 0.021037321537733078, + 0.0253921989351511, + 0.026811469346284866, + 0.015223911963403225, + 0.01781810075044632, + 0.025392737239599228, + 0.029368391260504723, + 0.02004241943359375, + 0.0070096515119075775, + 0.009426881559193134, + 0.009471338242292404, + 0.030669882893562317, + 0.007493745535612106, + 0.00835961103439331, + 0.01904113031923771, + 0.013249934650957584, + 0.008553988300263882, + 0.005881491582840681, + 0.01276326272636652, + 0.010077049024403095, + 0.007991671562194824, + 0.008269445970654488, + 0.013546692207455635, + 0.015272143296897411, + 0.013535465113818645, + 0.008896966464817524, + 0.009488549083471298, + 0.009361877106130123, + 0.011421586386859417, + 0.00861364509910345, + 0.01967610977590084, + 0.03829016909003258, + 0.00599694112315774, + 0.0058138370513916016, + 0.015355771407485008, + 0.01101923268288374, + 0.006530951242893934, + 0.005635402165353298, + 0.003156128339469433, + 0.0030087449122220278, + 0.0025929459370672703, + 0.0034968198742717505, + 0.0049197617918252945, + 0.006054183468222618, + 0.0032297931611537933, + 0.0037847706116735935, + 0.006858657579869032, + 0.005558378528803587, + 0.004027611576020718, + 0.0055816760286688805, + 0.0058869910426437855, + 0.005460771732032299, + 0.006282266229391098, + 0.005736384075134993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.023273516446352005 + } + ], + "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/0000060/group_9/index_tensor([8985, 8985])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_tensor([8985, 8985])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a212925775493c356c4797ae57799a7762719606 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_tensor([8985, 8985])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In vibrant Ghibli-inspired art, a dynamic low-angle tracking shot glides smoothly across a sun-drenched meadow, keeping pace with two European hares playfully chasing each other. Their forms are rendered in soft, painterly textures, characterized by warm browns and creams, their long, expressive ears twitching with delight. The entire scene is captured at a smooth, high frame rate, making every movement feel incredibly clear and vibrant. The lead hare hops gracefully, its body fully extended in a powerful, elegant arc against the bright, azure sky, while its companions joyfully weave through a lush, pastoral tapestry of red poppies, yellow buttercups, and other wildflowers that gently sway behind them. The landscape itself is a masterpiece of idyllic beauty, with rolling green hills cast long, soft shadows under the warm, golden sunlight. This painterly world, brimming with joyful energy and crisp detail in every leap and bound, evokes a profound sense of freedom and warm nostalgia.", + "seed": 23884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_8985_seed_23884_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": -0.027830651868134736, + "vlm_logprob": -0.027830651868134736, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.027830651868134736, + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08018248528242111 + ], + "opposite_logprob_tokens": [ + -2.5801825523376465 + ], + "logprob_reward": -0.08018248528242111, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08018248528242111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06761875003576279, + 0.10658500343561172, + 0.16514360904693604, + 0.0744674801826477, + 0.05150500684976578, + 0.09245017915964127, + 0.207557812333107, + 0.03657102957367897, + 0.13026738166809082, + 0.06783375889062881 + ], + "credit_map": [ + 0.012386965565383434, + 0.009761752560734749, + 0.008418882265686989, + 0.0073907505720853806, + 0.00604243203997612, + 0.0070185586810112, + 0.008554701693356037, + 0.008044708520174026, + 0.014881040900945663, + 0.011515618301928043, + 0.023786086589097977, + 0.01600792445242405, + 0.009515571407973766, + 0.006234515458345413, + 0.010832653380930424, + 0.013811593875288963, + 0.025808701291680336, + 0.025313980877399445, + 0.023862509056925774, + 0.016248458996415138, + 0.037706706672906876, + 0.010610698722302914, + 0.012069670483469963, + 0.01352288480848074, + 0.008999297395348549, + 0.011302148923277855, + 0.013132252730429173, + 0.01732613891363144, + 0.004532008897513151, + 0.0047719720751047134, + 0.006905749905854464, + 0.0074979038909077644, + 0.007350960746407509, + 0.008646287024021149, + 0.006730683613568544, + 0.005726434290409088, + 0.007968872785568237, + 0.006452946458011866, + 0.004237317945808172, + 0.004391505848616362, + 0.01047721691429615, + 0.013107753358781338, + 0.007999625988304615, + 0.015177950263023376, + 0.013662180863320827, + 0.008397079072892666, + 0.014571140520274639, + 0.00905723124742508, + 0.023784644901752472, + 0.024850554764270782, + 0.025786690413951874, + 0.02291019819676876, + 0.03096604160964489, + 0.03235452249646187, + 0.026774005964398384, + 0.02013116516172886, + 0.004619436804205179, + 0.0036310660652816296, + 0.004895669873803854, + 0.003837238298729062, + 0.0042795748449862, + 0.006526844576001167, + 0.004156868439167738, + 0.004624331835657358, + 0.013284096494317055, + 0.01202528364956379, + 0.018225260078907013, + 0.019973374903202057, + 0.014622092247009277, + 0.017209935933351517, + 0.018828632310032845, + 0.016098706051707268, + 0.008407669141888618, + 0.011321338824927807, + 0.008960123173892498, + 0.005883402191102505, + 0.008319037035107613, + 0.008399596437811852, + 0.010097053833305836, + 0.006445538252592087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.08018248528242111 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105948377400637 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.004105948377400637, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105948377400637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13671065866947174, + 0.0673358291387558, + 0.052359919995069504, + 0.07346058636903763, + 0.053974080830812454, + 0.12310009449720383, + 0.22732293605804443, + 0.023446006700396538, + 0.07467502355575562, + 0.16761484742164612 + ], + "credit_map": [ + 0.017194123938679695, + 0.01686250790953636, + 0.016661375761032104, + 0.014343024231493473, + 0.01816929504275322, + 0.016408873721957207, + 0.01872689090669155, + 0.018344558775424957, + 0.008254243060946465, + 0.008667214773595333, + 0.012117452919483185, + 0.006997763644903898, + 0.00907763559371233, + 0.006031871773302555, + 0.007725060917437077, + 0.008464585058391094, + 0.009691886603832245, + 0.011198841035366058, + 0.0072161308489739895, + 0.004992160480469465, + 0.005766575690358877, + 0.004543867893517017, + 0.004230376798659563, + 0.0047200811095535755, + 0.009306667372584343, + 0.014061038382351398, + 0.00941256619989872, + 0.015937482938170433, + 0.0068541658110916615, + 0.0059754676185548306, + 0.006159336306154728, + 0.005753862205892801, + 0.006767114624381065, + 0.006128612440079451, + 0.005806175991892815, + 0.008313625119626522, + 0.012584798038005829, + 0.004347019363194704, + 0.005043874494731426, + 0.00498286122456193, + 0.01600640080869198, + 0.018837161362171173, + 0.009721826761960983, + 0.02043391950428486, + 0.018230529502034187, + 0.010776109993457794, + 0.017753563821315765, + 0.011340590193867683, + 0.027236856520175934, + 0.03117690421640873, + 0.025152236223220825, + 0.024904202669858932, + 0.03802428022027016, + 0.029347384348511696, + 0.029895083978772163, + 0.02158597856760025, + 0.0031113645527511835, + 0.0026290344540029764, + 0.004434242844581604, + 0.0031364685855805874, + 0.0025884604547172785, + 0.0033490408677607775, + 0.0022558688651770353, + 0.0019415256101638079, + 0.007997807115316391, + 0.006580243352800608, + 0.010836186818778515, + 0.009002142585814, + 0.008801620453596115, + 0.008308741264045238, + 0.011355247348546982, + 0.011793041601777077, + 0.020093098282814026, + 0.02704007364809513, + 0.021156154572963715, + 0.01805577427148819, + 0.015116993337869644, + 0.017002005130052567, + 0.03221478313207626, + 0.016935965046286583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.004105948377400637 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249762907624245 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249762907624245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249762907624245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11576993018388748, + 0.1289099007844925, + 0.2824819087982178, + 0.12212501466274261, + 0.07865030318498611, + 0.05633776634931564, + 0.10876963287591934, + 0.027755513787269592, + 0.020476989448070526, + 0.05872305482625961 + ], + "credit_map": [ + 0.018225492909550667, + 0.017863804474473, + 0.014982124790549278, + 0.00924582127481699, + 0.01570221036672592, + 0.01370602659881115, + 0.013074340298771858, + 0.012970100156962872, + 0.016204413026571274, + 0.01214559655636549, + 0.03311998024582863, + 0.01780145801603794, + 0.009821461513638496, + 0.007591372821480036, + 0.016410687938332558, + 0.015814926475286484, + 0.04540642723441124, + 0.0391363799571991, + 0.05298030003905296, + 0.020997369661927223, + 0.0623120442032814, + 0.02138868346810341, + 0.018612276762723923, + 0.021648414433002472, + 0.015749972313642502, + 0.022107679396867752, + 0.016974348574876785, + 0.02306639589369297, + 0.009801531210541725, + 0.006352676078677177, + 0.013911359012126923, + 0.014161061495542526, + 0.011063759215176105, + 0.010486545972526073, + 0.00721435621380806, + 0.010049287229776382, + 0.020096220076084137, + 0.005738130770623684, + 0.007877806201577187, + 0.0061241933144629, + 0.007390523795038462, + 0.01005635317414999, + 0.006395719014108181, + 0.007230368442833424, + 0.008133490569889545, + 0.006692437920719385, + 0.0046281334944069386, + 0.0058107394725084305, + 0.011174321174621582, + 0.012655560858547688, + 0.011893277987837791, + 0.013632122427225113, + 0.016586624085903168, + 0.014050436206161976, + 0.0138864042237401, + 0.014890884049236774, + 0.0032817122992128134, + 0.0034312414936721325, + 0.004528462886810303, + 0.0026574961375445127, + 0.002207599114626646, + 0.005096739158034325, + 0.002976329531520605, + 0.0035759347956627607, + 0.0025182373356074095, + 0.0023678475990891457, + 0.0037877848371863365, + 0.002286262111738324, + 0.0021189593244343996, + 0.0018650515703484416, + 0.0025769046042114496, + 0.0029559433460235596, + 0.007301570847630501, + 0.010888923890888691, + 0.01070321910083294, + 0.00594334676861763, + 0.004964623600244522, + 0.00473391730338335, + 0.00880465842783451, + 0.005382797680795193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.005249762907624245 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634247951209545 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.02634247951209545, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634247951209545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18722674250602722, + 0.05822775885462761, + 0.07234199345111847, + 0.07329032570123672, + 0.025838760659098625, + 0.049623094499111176, + 0.0624593049287796, + 0.08203108608722687, + 0.0685904398560524, + 0.32037049531936646 + ], + "credit_map": [ + 0.016891401261091232, + 0.02382747456431389, + 0.017034800723195076, + 0.02022506296634674, + 0.02499929629266262, + 0.03392057120800018, + 0.02682475559413433, + 0.02350337617099285, + 0.005923449527472258, + 0.006898215971887112, + 0.007638097275048494, + 0.008211481384932995, + 0.0073252771981060505, + 0.006170977838337421, + 0.009388764388859272, + 0.00667149480432272, + 0.010190240107476711, + 0.010164455510675907, + 0.005916965194046497, + 0.005347600672394037, + 0.016770539805293083, + 0.007405004929751158, + 0.008623449131846428, + 0.007923733443021774, + 0.008240938186645508, + 0.011608485132455826, + 0.011116018518805504, + 0.01739962212741375, + 0.005924332421272993, + 0.00617159204557538, + 0.006741350516676903, + 0.006087991409003735, + 0.003276342758908868, + 0.003320221556350589, + 0.002153554232791066, + 0.0033365709241479635, + 0.005894802510738373, + 0.0025508494582027197, + 0.002856520237401128, + 0.0024498996790498495, + 0.007951815612614155, + 0.00889881607145071, + 0.003937504719942808, + 0.005443326197564602, + 0.009606577455997467, + 0.005812059156596661, + 0.003724234877154231, + 0.00424876157194376, + 0.008789436891674995, + 0.011702285148203373, + 0.007100121583789587, + 0.007225901819765568, + 0.006223070900887251, + 0.007831390015780926, + 0.005179163068532944, + 0.008407937362790108, + 0.009262836538255215, + 0.00966416671872139, + 0.014211630448698997, + 0.011684121564030647, + 0.008850494399666786, + 0.009933851659297943, + 0.009910701774060726, + 0.008513284847140312, + 0.007005890365689993, + 0.00604017311707139, + 0.011448798701167107, + 0.009251735173165798, + 0.006077246740460396, + 0.007511841133236885, + 0.007894212380051613, + 0.013360542245209217, + 0.04177243262529373, + 0.06304484605789185, + 0.04997647553682327, + 0.03301006183028221, + 0.026804393157362938, + 0.024736227467656136, + 0.04667193070054054, + 0.034354131668806076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.02634247951209545 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327258326113224 + ], + "opposite_logprob_tokens": [ + -3.7732725143432617 + ], + "logprob_reward": -0.02327258326113224, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327258326113224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08605532348155975, + 0.07226701825857162, + 0.045257169753313065, + 0.06835529208183289, + 0.07872407883405685, + 0.3550131022930145, + 0.08308492600917816, + 0.043805576860904694, + 0.11151546984910965, + 0.05592205002903938 + ], + "credit_map": [ + 0.01802748814225197, + 0.012397607788443565, + 0.012273372150957584, + 0.008219100534915924, + 0.007783007808029652, + 0.0077864546328783035, + 0.011819310486316681, + 0.007748978678137064, + 0.00856100395321846, + 0.008616180159151554, + 0.012991045601665974, + 0.010149030014872551, + 0.01182530540972948, + 0.004906532820314169, + 0.006743275094777346, + 0.008474643342196941, + 0.006891478784382343, + 0.007105913478881121, + 0.01003782544285059, + 0.005074845626950264, + 0.003572587389498949, + 0.0032292932737618685, + 0.004751498345285654, + 0.004593726713210344, + 0.009671247564256191, + 0.014931588433682919, + 0.008457468822598457, + 0.013033008202910423, + 0.006967378780245781, + 0.0043685645796358585, + 0.005208250135183334, + 0.005717786960303783, + 0.00997382402420044, + 0.009993602521717548, + 0.009098258800804615, + 0.013156196102499962, + 0.01350522879511118, + 0.005993904545903206, + 0.008905448950827122, + 0.008097609505057335, + 0.05840805917978287, + 0.08740774542093277, + 0.02787451446056366, + 0.046626899391412735, + 0.04529252275824547, + 0.02547535113990307, + 0.03744718059897423, + 0.02648083120584488, + 0.009803162887692451, + 0.010989119298756123, + 0.009539870545268059, + 0.011875061318278313, + 0.008322729729115963, + 0.012964076362550259, + 0.01084327232092619, + 0.008747638203203678, + 0.004469042178243399, + 0.004410112742334604, + 0.006129654124379158, + 0.005497634410858154, + 0.005828329361975193, + 0.006587789859622717, + 0.005075647495687008, + 0.005807369016110897, + 0.012243344448506832, + 0.011790738441050053, + 0.01669081673026085, + 0.018250202760100365, + 0.013943610712885857, + 0.013211210258305073, + 0.012601733207702637, + 0.012783816084265709, + 0.00684621836990118, + 0.008350302465260029, + 0.005797124933451414, + 0.005964292678982019, + 0.005437957588583231, + 0.00655438844114542, + 0.010700284503400326, + 0.006271480582654476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.02327258326113224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In vibrant Ghibli-inspired art, a dynamic low-angle tracking shot glides smoothly across a sun-drenched meadow, keeping pace with two European hares playfully chasing each other. Their forms are rendered in soft, painterly textures, characterized by warm browns and creams, their long, expressive ears twitching with delight. The entire scene is captured at a smooth, high frame rate, making every movement feel incredibly clear and vibrant. The lead hare hops gracefully, its body fully extended in a powerful, elegant arc against the bright, azure sky, while its companions joyfully weave through a lush, pastoral tapestry of red poppies, yellow buttercups, and other wildflowers that gently sway behind them. The landscape itself is a masterpiece of idyllic beauty, with rolling green hills cast long, soft shadows under the warm, golden sunlight. This painterly world, brimming with joyful energy and crisp detail in every leap and bound, evokes a profound sense of freedom and warm nostalgia.", + "seed": 23885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_8985_seed_23885_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": -0.018831091839820145, + "vlm_logprob": -0.018831091839820145, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.018831091839820145, + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044114600867033005 + ], + "opposite_logprob_tokens": [ + -3.169114589691162 + ], + "logprob_reward": -0.044114600867033005, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044114600867033005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14549139142036438, + 0.06849775463342667, + 0.2957378029823303, + 0.1442595273256302, + 0.030183710157871246, + 0.14074081182479858, + 0.04627816006541252, + 0.06829213351011276, + 0.030357932671904564, + 0.030160794034600258 + ], + "credit_map": [ + 0.01393156312406063, + 0.024418434128165245, + 0.013861618004739285, + 0.027201497927308083, + 0.022235918790102005, + 0.011715165339410305, + 0.019851017743349075, + 0.01227615587413311, + 0.005453633610159159, + 0.007274700794368982, + 0.01073884591460228, + 0.006837579887360334, + 0.006668453570455313, + 0.006853160448372364, + 0.010695857927203178, + 0.013975517824292183, + 0.03367419168353081, + 0.06779283285140991, + 0.028739212080836296, + 0.02076217532157898, + 0.033660273998975754, + 0.04915553703904152, + 0.030067116022109985, + 0.03188641741871834, + 0.02191525138914585, + 0.04373544827103615, + 0.009762270376086235, + 0.012215981259942055, + 0.012876077555119991, + 0.014724417589604855, + 0.015764927491545677, + 0.013265145942568779, + 0.0036086495965719223, + 0.008555829524993896, + 0.0024639887269586325, + 0.002072946634143591, + 0.0034904764033854008, + 0.003893472021445632, + 0.004068087320774794, + 0.002030257834121585, + 0.018904099240899086, + 0.04570276290178299, + 0.0312363188713789, + 0.011853388510644436, + 0.00967384222894907, + 0.009025239385664463, + 0.009460404515266418, + 0.004884741734713316, + 0.005996681284159422, + 0.01078734453767538, + 0.003923199605196714, + 0.004434532485902309, + 0.005756658036261797, + 0.004145288839936256, + 0.005717028398066759, + 0.005517418961971998, + 0.01339640747755766, + 0.026569511741399765, + 0.008656781166791916, + 0.004094105679541826, + 0.007136817555874586, + 0.002867736155167222, + 0.0024170756805688143, + 0.003153693163767457, + 0.0042394534684717655, + 0.004703288432210684, + 0.002760219154879451, + 0.002345355460420251, + 0.0033744722604751587, + 0.006557847373187542, + 0.00336631340906024, + 0.0030109798535704613, + 0.0025199875235557556, + 0.005164633505046368, + 0.0030036496464163065, + 0.006040241569280624, + 0.0026998957619071007, + 0.003701006295159459, + 0.0033025096636265516, + 0.003728868206962943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.044114600867033005 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028267446905374527 + ], + "opposite_logprob_tokens": [ + -5.877826690673828 + ], + "logprob_reward": -0.0028267446905374527, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028267446905374527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12089286744594574, + 0.03327864408493042, + 0.1475457400083542, + 0.03200360760092735, + 0.019565057009458542, + 0.1765066385269165, + 0.03449195250868797, + 0.3440096378326416, + 0.0647854283452034, + 0.026920421048998833 + ], + "credit_map": [ + 0.019027767702937126, + 0.03220409154891968, + 0.011590471491217613, + 0.0146968187764287, + 0.01119467243552208, + 0.008486932143568993, + 0.014981796033680439, + 0.00871032290160656, + 0.00330663425847888, + 0.003872218308970332, + 0.005245260428637266, + 0.0039554317481815815, + 0.0032355114817619324, + 0.004301765467971563, + 0.003993666730821133, + 0.005368157289922237, + 0.017843665555119514, + 0.03431488201022148, + 0.012924308888614178, + 0.010093619115650654, + 0.01569424383342266, + 0.02577631175518036, + 0.014358981512486935, + 0.016539720818400383, + 0.0032819847110658884, + 0.004651110153645277, + 0.003439538413658738, + 0.00484419334679842, + 0.0031961542554199696, + 0.002150763524696231, + 0.005393700674176216, + 0.005046162288635969, + 0.003047736594453454, + 0.006619288120418787, + 0.0014422073727473617, + 0.0012067012721672654, + 0.0012638928601518273, + 0.0027964080218225718, + 0.0016525998944416642, + 0.0015362235717475414, + 0.024757103994488716, + 0.059443339705467224, + 0.042575299739837646, + 0.014394450932741165, + 0.009950234554708004, + 0.011165246367454529, + 0.008312458172440529, + 0.0059085143730044365, + 0.0033815482165664434, + 0.0034931525588035583, + 0.0031574678141623735, + 0.0027507354971021414, + 0.008523128926753998, + 0.0030516686383634806, + 0.006525623146444559, + 0.00360862840898335, + 0.04633338749408722, + 0.13286450505256653, + 0.06162523105740547, + 0.017890935763716698, + 0.02448490634560585, + 0.0337430015206337, + 0.011938069015741348, + 0.015129596926271915, + 0.008433805778622627, + 0.008365158922970295, + 0.005727298557758331, + 0.006994228344410658, + 0.006375965662300587, + 0.011826361529529095, + 0.00824462529271841, + 0.008817982859909534, + 0.002781439572572708, + 0.0033779004588723183, + 0.0025265119038522243, + 0.0033781363163143396, + 0.0031779299024492502, + 0.005474308039993048, + 0.0029705322813242674, + 0.0032336628064513206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.0028267446905374527 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191023599356413 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.003191023599356413, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191023599356413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628289759159088, + 0.06806185841560364, + 0.34979379177093506, + 0.020101383328437805, + 0.032267581671476364, + 0.02484913542866707, + 0.029284970834851265, + 0.163564994931221, + 0.02439633384346962, + 0.024850981310009956 + ], + "credit_map": [ + 0.03918680548667908, + 0.06518848240375519, + 0.025653965771198273, + 0.0350753553211689, + 0.02505628950893879, + 0.017722712829709053, + 0.03311963006854057, + 0.021825747564435005, + 0.0061943563632667065, + 0.008112008683383465, + 0.01156540960073471, + 0.00987872015684843, + 0.008008013479411602, + 0.005516393575817347, + 0.007179347798228264, + 0.011607605032622814, + 0.04285908862948418, + 0.08207671344280243, + 0.03404436632990837, + 0.025768382474780083, + 0.03695863112807274, + 0.05421193316578865, + 0.03546583652496338, + 0.038408827036619186, + 0.003160650609061122, + 0.004778687842190266, + 0.0016148419817909598, + 0.002232635859400034, + 0.0018920967122539878, + 0.001279862248338759, + 0.002695626113563776, + 0.0024469823110848665, + 0.004442392382770777, + 0.010029340162873268, + 0.00210962095297873, + 0.00200481154024601, + 0.0032385161612182856, + 0.00471768993884325, + 0.003409281838685274, + 0.0023159291595220566, + 0.0036521146539598703, + 0.00576996523886919, + 0.0029848329722881317, + 0.0024799618404358625, + 0.0022572793532162905, + 0.0015315711498260498, + 0.003631848841905594, + 0.0025415620766580105, + 0.0037752515636384487, + 0.004807701800018549, + 0.002495395950973034, + 0.0029431902803480625, + 0.004762479569762945, + 0.0025228294543921947, + 0.005382002331316471, + 0.0025961215142160654, + 0.02355448342859745, + 0.06412846595048904, + 0.028008844703435898, + 0.008120578713715076, + 0.01275833323597908, + 0.014696571975946426, + 0.00554331811144948, + 0.006754405330866575, + 0.004582565743476152, + 0.0026827892288565636, + 0.001961753936484456, + 0.0017669593216851354, + 0.0026544537395238876, + 0.006965228356420994, + 0.0018900064751505852, + 0.001892576925456524, + 0.003031147876754403, + 0.005028044804930687, + 0.002127910265699029, + 0.0041066305711865425, + 0.0021187416277825832, + 0.0036159930750727654, + 0.0026636491529643536, + 0.0021588648669421673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.003191023599356413 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942889489233494 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942889489233494, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942889489233494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11275605857372284, + 0.06753896921873093, + 0.06457508355379105, + 0.10642720013856888, + 0.0819527730345726, + 0.14600953459739685, + 0.1486143320798874, + 0.19804422557353973, + 0.0391167588531971, + 0.03496507182717323 + ], + "credit_map": [ + 0.014835589565336704, + 0.025892488658428192, + 0.01688072644174099, + 0.01384925190359354, + 0.011441733688116074, + 0.00779902096837759, + 0.01040536630898714, + 0.011651880107820034, + 0.007515126373618841, + 0.012413782998919487, + 0.007392916828393936, + 0.005707288160920143, + 0.007330894935876131, + 0.011378404684364796, + 0.005686793010681868, + 0.010113759897649288, + 0.00584882078692317, + 0.008700043894350529, + 0.008276169188320637, + 0.009387068450450897, + 0.012017575092613697, + 0.005297488998621702, + 0.009127753786742687, + 0.005920164752751589, + 0.020805181935429573, + 0.03090442717075348, + 0.008240209892392159, + 0.011175942607223988, + 0.007898877374827862, + 0.007040478754788637, + 0.00983365811407566, + 0.010528422892093658, + 0.012767795473337173, + 0.02686513401567936, + 0.006916190963238478, + 0.00589753407984972, + 0.007155807688832283, + 0.00977384578436613, + 0.006270607467740774, + 0.006305857561528683, + 0.017958562821149826, + 0.022107744589447975, + 0.031080612912774086, + 0.022250661626458168, + 0.0132682453840971, + 0.011381931602954865, + 0.01452577207237482, + 0.013435999862849712, + 0.01574963703751564, + 0.03067370504140854, + 0.02384897507727146, + 0.0174639243632555, + 0.014267810620367527, + 0.014297381043434143, + 0.01730421744287014, + 0.015008683316409588, + 0.035294681787490845, + 0.0767061710357666, + 0.03313883766531944, + 0.010573532432317734, + 0.018111662939190865, + 0.00865069217979908, + 0.007103170268237591, + 0.008465473540127277, + 0.005236127879470587, + 0.006084901746362448, + 0.004541448783129454, + 0.004975340329110622, + 0.004927980247884989, + 0.005324636120349169, + 0.004742157645523548, + 0.0032841672655194998, + 0.003815164091065526, + 0.00485703069716692, + 0.003337370464578271, + 0.0046815103851258755, + 0.003739685518667102, + 0.005445257760584354, + 0.005064674187451601, + 0.004024380352348089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.005942889489233494 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808020055294037 + ], + "opposite_logprob_tokens": [ + -3.2880802154541016 + ], + "logprob_reward": -0.03808020055294037, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808020055294037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09033353626728058, + 0.049807533621788025, + 0.1616501659154892, + 0.06225382164120674, + 0.053329065442085266, + 0.026752064004540443, + 0.07152623683214188, + 0.32834485173225403, + 0.13444778323173523, + 0.021554943174123764 + ], + "credit_map": [ + 0.011202951893210411, + 0.012628012336790562, + 0.009640611708164215, + 0.011871326714754105, + 0.016088923439383507, + 0.007246278692036867, + 0.013855245895683765, + 0.007800184655934572, + 0.0045453691855072975, + 0.006336906924843788, + 0.005775217432528734, + 0.007948528975248337, + 0.006151977926492691, + 0.003982732072472572, + 0.006232462357729673, + 0.008834338746964931, + 0.01922106370329857, + 0.036675747483968735, + 0.015463307499885559, + 0.012307864613831043, + 0.016345899552106857, + 0.026177076622843742, + 0.017096105962991714, + 0.018363110721111298, + 0.0055935378186404705, + 0.006442884914577007, + 0.006290648598223925, + 0.008695494383573532, + 0.0071275015361607075, + 0.004104359075427055, + 0.012964808382093906, + 0.011034587398171425, + 0.0073184724897146225, + 0.01750989630818367, + 0.0031711338087916374, + 0.002703395439311862, + 0.005072603467851877, + 0.008511742576956749, + 0.005144961178302765, + 0.003896859474480152, + 0.0027745410334318876, + 0.006201472599059343, + 0.006851712241768837, + 0.0026376808527857065, + 0.003201153827831149, + 0.0016440772451460361, + 0.001999452244490385, + 0.0014419739600270987, + 0.007064808160066605, + 0.008355192840099335, + 0.0069443886168301105, + 0.0058051543310284615, + 0.018403945490717888, + 0.007238528225570917, + 0.011223924346268177, + 0.0064902943558990955, + 0.044578567147254944, + 0.12631915509700775, + 0.05991339683532715, + 0.017001722007989883, + 0.024824250489473343, + 0.029293647035956383, + 0.01207231916487217, + 0.014341787435114384, + 0.013609451241791248, + 0.021174253895878792, + 0.013191556558012962, + 0.017452863976359367, + 0.01574326492846012, + 0.014909563586115837, + 0.020900584757328033, + 0.017466241493821144, + 0.0023546670563519, + 0.0030354545451700687, + 0.0017928571905940771, + 0.004215533845126629, + 0.0017608464695513248, + 0.0032963112462311983, + 0.003061748342588544, + 0.002037524711340666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.03808020055294037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In vibrant Ghibli-inspired art, a dynamic low-angle tracking shot glides smoothly across a sun-drenched meadow, keeping pace with two European hares playfully chasing each other. Their forms are rendered in soft, painterly textures, characterized by warm browns and creams, their long, expressive ears twitching with delight. The entire scene is captured at a smooth, high frame rate, making every movement feel incredibly clear and vibrant. The lead hare hops gracefully, its body fully extended in a powerful, elegant arc against the bright, azure sky, while its companions joyfully weave through a lush, pastoral tapestry of red poppies, yellow buttercups, and other wildflowers that gently sway behind them. The landscape itself is a masterpiece of idyllic beauty, with rolling green hills cast long, soft shadows under the warm, golden sunlight. This painterly world, brimming with joyful energy and crisp detail in every leap and bound, evokes a profound sense of freedom and warm nostalgia.", + "seed": 23886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_8985_seed_23886_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": -0.03257561251521111, + "vlm_logprob": -0.03257561251521111, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03257561251521111, + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06309987604618073 + ], + "opposite_logprob_tokens": [ + -2.8130998611450195 + ], + "logprob_reward": -0.06309987604618073, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06309987604618073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07801275700330734, + 0.055661749094724655, + 0.12125878781080246, + 0.1428033709526062, + 0.08807246387004852, + 0.19449827075004578, + 0.026547322049736977, + 0.09276105463504791, + 0.178282231092453, + 0.022101936861872673 + ], + "credit_map": [ + 0.008936071768403053, + 0.013384953141212463, + 0.00751112774014473, + 0.010366815142333508, + 0.016249418258666992, + 0.006778137758374214, + 0.006878257263451815, + 0.007907983846962452, + 0.00814328994601965, + 0.012781682424247265, + 0.007858009077608585, + 0.005973318591713905, + 0.006392599083483219, + 0.004034710116684437, + 0.005653861910104752, + 0.004824278876185417, + 0.016559498384594917, + 0.022368986159563065, + 0.01334309671074152, + 0.016683246940374374, + 0.020319903269410133, + 0.01198067981749773, + 0.01102271769195795, + 0.008980678394436836, + 0.021527845412492752, + 0.02449079602956772, + 0.013356706127524376, + 0.02313200943171978, + 0.02357003279030323, + 0.013598824851214886, + 0.008601763285696507, + 0.014525393955409527, + 0.01152811013162136, + 0.010320975445210934, + 0.009331815876066685, + 0.013247179798781872, + 0.015782294794917107, + 0.010081800632178783, + 0.006960552651435137, + 0.010819735005497932, + 0.026047833263874054, + 0.02132619172334671, + 0.019118180498480797, + 0.02825191058218479, + 0.023164333775639534, + 0.03056500107049942, + 0.022892924025654793, + 0.023131918162107468, + 0.004196684341877699, + 0.003901997348293662, + 0.0019956447649747133, + 0.00357679626904428, + 0.0055687762796878815, + 0.0034819876309484243, + 0.0020918939262628555, + 0.0017335437005385756, + 0.011958103626966476, + 0.010600686073303223, + 0.008371458388864994, + 0.011761716566979885, + 0.02804216556251049, + 0.005596932955086231, + 0.01002185232937336, + 0.006408146116882563, + 0.04473477974534035, + 0.04674568399786949, + 0.008298376575112343, + 0.0109037384390831, + 0.020828217267990112, + 0.017169097438454628, + 0.017665334045886993, + 0.01193702220916748, + 0.0018100314773619175, + 0.0022598551586270332, + 0.002600237727165222, + 0.0030845936853438616, + 0.0019090618006885052, + 0.003017580369487405, + 0.004291566088795662, + 0.0031290126498788595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.06309987604618073 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647525027394295 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647525027394295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647525027394295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15678755939006805, + 0.08234088867902756, + 0.09420201182365417, + 0.3145535886287689, + 0.1053043082356453, + 0.04867444187402725, + 0.04036426544189453, + 0.0674438625574112, + 0.05703195184469223, + 0.03329715505242348 + ], + "credit_map": [ + 0.013941194862127304, + 0.03224077820777893, + 0.011563451960682869, + 0.018258625641465187, + 0.038336899131536484, + 0.01060804445296526, + 0.01663309335708618, + 0.015205473639070988, + 0.013034137897193432, + 0.011589013040065765, + 0.00766502832993865, + 0.0059575107879936695, + 0.011059045791625977, + 0.008461717516183853, + 0.012847297824919224, + 0.011727139353752136, + 0.012063521891832352, + 0.01679612323641777, + 0.014187926426529884, + 0.013734155334532261, + 0.008613299578428268, + 0.008103919215500355, + 0.009155778214335442, + 0.011547290720045567, + 0.04012352600693703, + 0.020760484039783478, + 0.05401964858174324, + 0.0355551578104496, + 0.03385733440518379, + 0.04028758779168129, + 0.03133619949221611, + 0.05861363187432289, + 0.013660716824233532, + 0.010227044112980366, + 0.017239965498447418, + 0.011585565283894539, + 0.018335510045289993, + 0.013700180687010288, + 0.00758079718798399, + 0.012974529527127743, + 0.005211806856095791, + 0.004746811930090189, + 0.00708413403481245, + 0.007023348938673735, + 0.0055891647934913635, + 0.007678414694964886, + 0.0049234842881560326, + 0.006417274475097656, + 0.0056977057829499245, + 0.004777913447469473, + 0.005317441187798977, + 0.004660320468246937, + 0.007681462913751602, + 0.005062034819275141, + 0.003641054965555668, + 0.003526329528540373, + 0.010135678574442863, + 0.007621545810252428, + 0.008206160739064217, + 0.008736951276659966, + 0.016028571873903275, + 0.004814703483134508, + 0.007394515443593264, + 0.004505737219005823, + 0.013342204503715038, + 0.013896861113607883, + 0.0038066578563302755, + 0.004921718500554562, + 0.00597924180328846, + 0.003643697826191783, + 0.006960968021303415, + 0.004480601754039526, + 0.0028125084936618805, + 0.0035381291527301073, + 0.005579208955168724, + 0.003463673172518611, + 0.0030836304649710655, + 0.0055312504991889, + 0.005598022602498531, + 0.0036907331086695194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.004647525027394295 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17858873307704926, + 0.08913274109363556, + 0.1152416244149208, + 0.31249886751174927, + 0.05069490522146225, + 0.04591909423470497, + 0.020674021914601326, + 0.04589441046118736, + 0.08765759319067001, + 0.05369795858860016 + ], + "credit_map": [ + 0.01687864400446415, + 0.029526229947805405, + 0.01761271245777607, + 0.020091328769922256, + 0.031486328691244125, + 0.012353133410215378, + 0.025068892166018486, + 0.025571472942829132, + 0.015059512108564377, + 0.012677313759922981, + 0.016297737136483192, + 0.009504176676273346, + 0.012868660502135754, + 0.005678198765963316, + 0.009064937941730022, + 0.00798221118748188, + 0.015512548387050629, + 0.02742842026054859, + 0.009461452253162861, + 0.01575060747563839, + 0.015238003805279732, + 0.011147045530378819, + 0.009980307891964912, + 0.010723238810896873, + 0.0425567552447319, + 0.041786372661590576, + 0.037666141986846924, + 0.0527895912528038, + 0.05698644369840622, + 0.027833348140120506, + 0.02316259779036045, + 0.02971760369837284, + 0.006650249939411879, + 0.006400663405656815, + 0.0062184943817555904, + 0.006917421240359545, + 0.009770267643034458, + 0.005488918628543615, + 0.0035030946601182222, + 0.005745795089751482, + 0.005358108319342136, + 0.005942243151366711, + 0.0064099133014678955, + 0.005758268292993307, + 0.0061690895818173885, + 0.0051594628021121025, + 0.006527573801577091, + 0.0045944335870444775, + 0.0028189285658299923, + 0.0021129942033439875, + 0.002370760077610612, + 0.0019930992275476456, + 0.005383589770644903, + 0.002484765835106373, + 0.0017997929826378822, + 0.0017100912518799305, + 0.006142130121588707, + 0.0077354260720312595, + 0.003680487396195531, + 0.007307978812605143, + 0.01039411686360836, + 0.0035518419463187456, + 0.004480942152440548, + 0.0026014891918748617, + 0.019562434405088425, + 0.019583119079470634, + 0.005625660065561533, + 0.0077854604460299015, + 0.00701814703643322, + 0.011376024223864079, + 0.011956301517784595, + 0.004750446882098913, + 0.00350950681604445, + 0.004805341362953186, + 0.004721260163933039, + 0.008061687462031841, + 0.003713997546583414, + 0.007569880690425634, + 0.012498165480792522, + 0.008818117901682854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012535471469163895 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057681791484356 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.02057681791484356, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057681791484356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12831957638263702, + 0.05398794263601303, + 0.06511618942022324, + 0.46124351024627686, + 0.11406607925891876, + 0.025009773671627045, + 0.02335512824356556, + 0.01935999095439911, + 0.04617711156606674, + 0.06336471438407898 + ], + "credit_map": [ + 0.026766519993543625, + 0.011834349483251572, + 0.01049702800810337, + 0.014497620984911919, + 0.026469089090824127, + 0.012501448392868042, + 0.012973740696907043, + 0.012779772281646729, + 0.00870454590767622, + 0.008472918532788754, + 0.006983194965869188, + 0.0042373305186629295, + 0.006966008804738522, + 0.006082294043153524, + 0.006169175263494253, + 0.006372477393597364, + 0.0074983215890824795, + 0.009456870146095753, + 0.009245472960174084, + 0.00976645853370428, + 0.005396893713623285, + 0.007407008670270443, + 0.007321443874388933, + 0.00902372132986784, + 0.056335169821977615, + 0.024296091869473457, + 0.07256362587213516, + 0.05310327932238579, + 0.03983078896999359, + 0.06953790783882141, + 0.04628302901983261, + 0.09929360449314117, + 0.014554445631802082, + 0.014240618795156479, + 0.016396697610616684, + 0.01657908782362938, + 0.018729394301772118, + 0.013352317735552788, + 0.008974823169410229, + 0.011238695122301579, + 0.0024618911556899548, + 0.002541125053539872, + 0.006388024426996708, + 0.003988529555499554, + 0.0020481611136347055, + 0.0022189500741660595, + 0.002594396471977234, + 0.002768694655969739, + 0.003081070724874735, + 0.002918150508776307, + 0.0027372476179152727, + 0.0027560866437852383, + 0.003072963794693351, + 0.0035778339952230453, + 0.002381692174822092, + 0.0028300816193223, + 0.0027324918191879988, + 0.0025298960972577333, + 0.0019999162759631872, + 0.003458156017586589, + 0.003945131320506334, + 0.001640629954636097, + 0.0016773209208622575, + 0.001376449130475521, + 0.015182185918092728, + 0.015881765633821487, + 0.002104716142639518, + 0.002776762470602989, + 0.002650314010679722, + 0.0029716540593653917, + 0.0028805129695683718, + 0.0017291989643126726, + 0.005409142468124628, + 0.007535055745393038, + 0.004468034487217665, + 0.005321469157934189, + 0.008296193554997444, + 0.01968216337263584, + 0.006757176015526056, + 0.005895477719604969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.02057681791484356 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06201837211847305 + ], + "opposite_logprob_tokens": [ + -2.812018394470215 + ], + "logprob_reward": -0.06201837211847305, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06201837211847305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1169353500008583, + 0.07841174304485321, + 0.17279337346553802, + 0.11096996814012527, + 0.1245054230093956, + 0.106700599193573, + 0.036840200424194336, + 0.07191570103168488, + 0.11673638224601746, + 0.06419118493795395 + ], + "credit_map": [ + 0.01806749403476715, + 0.01880773901939392, + 0.008507234044373035, + 0.014713626354932785, + 0.03477059677243233, + 0.006986611522734165, + 0.00738806976005435, + 0.007693983148783445, + 0.010330484248697758, + 0.018247688189148903, + 0.007023794110864401, + 0.010461129248142242, + 0.008643146604299545, + 0.005156602244824171, + 0.009339396841824055, + 0.009209503419697285, + 0.022382168099284172, + 0.02488558180630207, + 0.028306856751441956, + 0.03258268162608147, + 0.014539078809320927, + 0.012009356170892715, + 0.01918719708919525, + 0.018900442868471146, + 0.01283316221088171, + 0.0152592733502388, + 0.010951594449579716, + 0.023002423346042633, + 0.01676175557076931, + 0.012043669819831848, + 0.011143837124109268, + 0.00897425040602684, + 0.017484789714217186, + 0.016027165576815605, + 0.014225056394934654, + 0.014586377888917923, + 0.025205526500940323, + 0.012920173816382885, + 0.010566653683781624, + 0.013489686883985996, + 0.011446096003055573, + 0.011026103049516678, + 0.0077331410720944405, + 0.015332459472119808, + 0.014803033322095871, + 0.014084975235164165, + 0.019929852336645126, + 0.012344940565526485, + 0.005334422457963228, + 0.004416962154209614, + 0.0033360684756189585, + 0.004461129195988178, + 0.008244717493653297, + 0.005281850229948759, + 0.0028530105482786894, + 0.0029120370745658875, + 0.011273609474301338, + 0.014341726899147034, + 0.009429844096302986, + 0.009387013502418995, + 0.008964388631284237, + 0.005423139315098524, + 0.007224820554256439, + 0.00587115902453661, + 0.030181320384144783, + 0.02968776598572731, + 0.007211138028651476, + 0.009957350790500641, + 0.010778567753732204, + 0.007147584110498428, + 0.014606609009206295, + 0.007166053168475628, + 0.004050455056130886, + 0.006439251825213432, + 0.006328259594738483, + 0.00875389389693737, + 0.0046394760720431805, + 0.008724760264158249, + 0.015408030711114407, + 0.009847057983279228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.06201837211847305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In vibrant Ghibli-inspired art, a dynamic low-angle tracking shot glides smoothly across a sun-drenched meadow, keeping pace with two European hares playfully chasing each other. Their forms are rendered in soft, painterly textures, characterized by warm browns and creams, their long, expressive ears twitching with delight. The entire scene is captured at a smooth, high frame rate, making every movement feel incredibly clear and vibrant. The lead hare hops gracefully, its body fully extended in a powerful, elegant arc against the bright, azure sky, while its companions joyfully weave through a lush, pastoral tapestry of red poppies, yellow buttercups, and other wildflowers that gently sway behind them. The landscape itself is a masterpiece of idyllic beauty, with rolling green hills cast long, soft shadows under the warm, golden sunlight. This painterly world, brimming with joyful energy and crisp detail in every leap and bound, evokes a profound sense of freedom and warm nostalgia.", + "seed": 23887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000060/group_9/index_8985_seed_23887_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": -0.07290457971394063, + "vlm_logprob": -0.07290457971394063, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07290457971394063, + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0808902382850647 + ], + "opposite_logprob_tokens": [ + -2.58089017868042 + ], + "logprob_reward": -0.0808902382850647, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0808902382850647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23505936563014984, + 0.07524807751178741, + 0.11382085084915161, + 0.05472991243004799, + 0.18459442257881165, + 0.10840251296758652, + 0.08083081245422363, + 0.06413592398166656, + 0.03242616355419159, + 0.05075196549296379 + ], + "credit_map": [ + 0.02317364700138569, + 0.020814325660467148, + 0.04390616714954376, + 0.02500670589506626, + 0.019500570371747017, + 0.0456526018679142, + 0.01926029659807682, + 0.03774505853652954, + 0.007421054411679506, + 0.009535577148199081, + 0.011934559792280197, + 0.013164118863642216, + 0.007777179125696421, + 0.006330982781946659, + 0.008238758891820908, + 0.010845839977264404, + 0.011077076196670532, + 0.012267230078577995, + 0.02240990474820137, + 0.013302210718393326, + 0.015792379155755043, + 0.015612470917403698, + 0.010191979818046093, + 0.013167602010071278, + 0.004292324185371399, + 0.00632136594504118, + 0.008019349537789822, + 0.007565398700535297, + 0.008054317906498909, + 0.007386005483567715, + 0.006440760102123022, + 0.006650392897427082, + 0.012284621596336365, + 0.01682373881340027, + 0.024605752900242805, + 0.028086192905902863, + 0.0189792662858963, + 0.03356606513261795, + 0.026252862066030502, + 0.023995928466320038, + 0.009142857044935226, + 0.01377728022634983, + 0.014099959284067154, + 0.012085367925465107, + 0.01912510022521019, + 0.017066478729248047, + 0.011196183040738106, + 0.011909286491572857, + 0.005110334139317274, + 0.008739011362195015, + 0.011603588238358498, + 0.013748107478022575, + 0.00987597182393074, + 0.007383659482002258, + 0.012865062803030014, + 0.011505072936415672, + 0.003279848024249077, + 0.0037812686059623957, + 0.01686706766486168, + 0.018574830144643784, + 0.00647346954792738, + 0.003387044183909893, + 0.005839736666530371, + 0.005932657979428768, + 0.00163756194524467, + 0.002086603082716465, + 0.005396317224949598, + 0.003966126125305891, + 0.0023572947829961777, + 0.004779279697686434, + 0.007805942092090845, + 0.004397040698677301, + 0.0031950350385159254, + 0.003947763703763485, + 0.00886318739503622, + 0.006259526126086712, + 0.004287535324692726, + 0.0059205046854913235, + 0.007835379801690578, + 0.010443035513162613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0808902382850647 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792858734726906 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792858734726906, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792858734726906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31564098596572876, + 0.0931265726685524, + 0.16735470294952393, + 0.03967570886015892, + 0.09471455216407776, + 0.09835704416036606, + 0.03136199712753296, + 0.08404050022363663, + 0.03410060331225395, + 0.04162722826004028 + ], + "credit_map": [ + 0.031105203554034233, + 0.03379301354289055, + 0.043680720031261444, + 0.02813779003918171, + 0.027330642566084862, + 0.06382350623607635, + 0.02703573741018772, + 0.060734424740076065, + 0.010084169916808605, + 0.013601580634713173, + 0.010783836245536804, + 0.015144814737141132, + 0.012941497378051281, + 0.009613454341888428, + 0.01100304163992405, + 0.009954183362424374, + 0.01719970628619194, + 0.020026367157697678, + 0.031914446502923965, + 0.019026411697268486, + 0.024349192157387733, + 0.02215314842760563, + 0.01675337366759777, + 0.015932094305753708, + 0.004898798651993275, + 0.004051746800541878, + 0.006239510141313076, + 0.006229887250810862, + 0.004418495576828718, + 0.0037574677262455225, + 0.005079975817352533, + 0.004999835509806871, + 0.00685315765440464, + 0.01080153789371252, + 0.012486999854445457, + 0.014960112981498241, + 0.010703153908252716, + 0.018068823963403702, + 0.010416967794299126, + 0.010423812083899975, + 0.008783825673162937, + 0.008689139038324356, + 0.013985910452902317, + 0.011638500727713108, + 0.012635236606001854, + 0.013473067432641983, + 0.016986126080155373, + 0.012165253050625324, + 0.002224414376541972, + 0.0029684011824429035, + 0.004266232717782259, + 0.004077895078808069, + 0.003947373013943434, + 0.00508573604747653, + 0.004869819153100252, + 0.003922130446881056, + 0.0036955701652914286, + 0.004397377837449312, + 0.024619093164801598, + 0.027884192764759064, + 0.008173909038305283, + 0.003486617701128125, + 0.007435380015522242, + 0.004348368849605322, + 0.002291710115969181, + 0.0032885102555155754, + 0.0068842750042676926, + 0.005084328353404999, + 0.002731196116656065, + 0.004947674926370382, + 0.0045190900564193726, + 0.00435382267460227, + 0.002656226744875312, + 0.004200205206871033, + 0.01049291342496872, + 0.0058798035606741905, + 0.002996449125930667, + 0.0045827352441847324, + 0.004746865015476942, + 0.006072039715945721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.009792858734726906 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525154173374176 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.00525154173374176, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525154173374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26116418838500977, + 0.06497599929571152, + 0.3012704849243164, + 0.15193849802017212, + 0.04223835840821266, + 0.06350374221801758, + 0.021460246294736862, + 0.01966490037739277, + 0.022839801385998726, + 0.05094368755817413 + ], + "credit_map": [ + 0.028353817760944366, + 0.027051184326410294, + 0.03386417031288147, + 0.02968745119869709, + 0.02230214700102806, + 0.05276377499103546, + 0.026794159784913063, + 0.040347494184970856, + 0.00688898004591465, + 0.012006537988781929, + 0.006521264091134071, + 0.010020721703767776, + 0.009936517104506493, + 0.006613082718104124, + 0.00783617701381445, + 0.005152728408575058, + 0.02660275809466839, + 0.033419858664274216, + 0.05998501181602478, + 0.0335480198264122, + 0.048902664333581924, + 0.03825389966368675, + 0.026577994227409363, + 0.033980317413806915, + 0.01640978269279003, + 0.016240349039435387, + 0.025555623695254326, + 0.021939581260085106, + 0.01842786744236946, + 0.019039027392864227, + 0.02030467614531517, + 0.014021600596606731, + 0.0023858684580773115, + 0.004246925935149193, + 0.005524822976440191, + 0.009729750454425812, + 0.004252966493368149, + 0.0074178180657327175, + 0.005015197210013866, + 0.0036650148686021566, + 0.006075817160308361, + 0.0068081882782280445, + 0.00970219075679779, + 0.007075686939060688, + 0.007617343682795763, + 0.007907774299383163, + 0.009374825283885002, + 0.00894191861152649, + 0.0015539367450401187, + 0.0015402999706566334, + 0.0029376826714724302, + 0.0034282817505300045, + 0.00329322530888021, + 0.002120544668287039, + 0.0038024301175028086, + 0.002783846575766802, + 0.0012147828238084912, + 0.0014157005352899432, + 0.005120510701090097, + 0.005662089213728905, + 0.0017356086755171418, + 0.0008840038790367544, + 0.0019616391509771347, + 0.0016705687157809734, + 0.0014231083914637566, + 0.00214327871799469, + 0.0033397532533854246, + 0.003283755388110876, + 0.0020527674350887537, + 0.003164243418723345, + 0.00418593967333436, + 0.0032469553407281637, + 0.0021880825515836477, + 0.0036052337381988764, + 0.009906318970024586, + 0.007472314406186342, + 0.0035231865476816893, + 0.0073085869662463665, + 0.006960379891097546, + 0.00997958518564701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.00525154173374176 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060317635536194 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.043060317635536194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060317635536194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3238389194011688, + 0.10472606122493744, + 0.15040741860866547, + 0.06117621064186096, + 0.040927693247795105, + 0.19154459238052368, + 0.026126548647880554, + 0.030495300889015198, + 0.03654738515615463, + 0.03420991823077202 + ], + "credit_map": [ + 0.03277882933616638, + 0.03401206061244011, + 0.04037241265177727, + 0.02670932188630104, + 0.02552320994436741, + 0.06578047573566437, + 0.030324487015604973, + 0.06833811849355698, + 0.011706423945724964, + 0.018930044025182724, + 0.009656746871769428, + 0.015652405098080635, + 0.01613869145512581, + 0.011173726990818977, + 0.012968625873327255, + 0.008499390445649624, + 0.014746918343007565, + 0.017108594998717308, + 0.030060403048992157, + 0.0179037693887949, + 0.022918056696653366, + 0.019143447279930115, + 0.014089439064264297, + 0.014436779543757439, + 0.007639634422957897, + 0.007706206291913986, + 0.01004478707909584, + 0.00799119845032692, + 0.007053499575704336, + 0.008130460977554321, + 0.007209130562841892, + 0.005401287227869034, + 0.002759448252618313, + 0.004590822383761406, + 0.005254935473203659, + 0.0044981203973293304, + 0.004835613537579775, + 0.007532613817602396, + 0.0043279253877699375, + 0.0071282112039625645, + 0.01895250752568245, + 0.025919627398252487, + 0.028545286506414413, + 0.019252276048064232, + 0.027650201693177223, + 0.02448628842830658, + 0.024584537371993065, + 0.022153843194246292, + 0.001909041777253151, + 0.0023565387818962336, + 0.0035387331154197454, + 0.002929726615548134, + 0.0030662689823657274, + 0.004958069883286953, + 0.00415634922683239, + 0.003211819799616933, + 0.0021266546100378036, + 0.0030474525410681963, + 0.006785108707845211, + 0.006770540028810501, + 0.003080641385167837, + 0.0016885052900761366, + 0.0038054215256124735, + 0.0031909735407680273, + 0.002192800398916006, + 0.0033897554967552423, + 0.00721513107419014, + 0.005739997606724501, + 0.003190515097230673, + 0.0051289768889546394, + 0.005637940485030413, + 0.004052266012877226, + 0.0019751647487282753, + 0.003078697482123971, + 0.008571991696953773, + 0.005366245750337839, + 0.002691387664526701, + 0.0028099766932427883, + 0.004225574433803558, + 0.0054908753372728825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.043060317635536194 + }, + { + "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\": \"Are two European hares visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the art style vibrant and Ghibli-inspired with soft, painterly textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hares colored in warm browns and creams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a sun-drenched meadow with red poppies, yellow buttercups, and wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a dynamic low-angle tracking shot gliding smoothly from behind the hares?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22552794218063354 + ], + "opposite_logprob_tokens": [ + -1.6005280017852783 + ], + "logprob_reward": -0.22552794218063354, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22552794218063354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11298130452632904, + 0.17264476418495178, + 0.10875443369150162, + 0.08999793231487274, + 0.13523608446121216, + 0.04268293455243111, + 0.03347364813089371, + 0.16851367056369781, + 0.094648577272892, + 0.041066646575927734 + ], + "credit_map": [ + 0.01420485321432352, + 0.012928423471748829, + 0.01530896034091711, + 0.014269945211708546, + 0.009013202041387558, + 0.015645546838641167, + 0.01084695104509592, + 0.020763425156474113, + 0.019243789836764336, + 0.03316484019160271, + 0.01751355640590191, + 0.02697165682911873, + 0.02416808344423771, + 0.015048705041408539, + 0.020016858354210854, + 0.01651727966964245, + 0.011304019019007683, + 0.012196309864521027, + 0.01755499839782715, + 0.012110354378819466, + 0.013535224832594395, + 0.01938598044216633, + 0.010244794189929962, + 0.012422758154571056, + 0.012269283644855022, + 0.008792977780103683, + 0.017709117382764816, + 0.012264234945178032, + 0.008046752773225307, + 0.008818800561130047, + 0.012250043451786041, + 0.009846730157732964, + 0.006794771179556847, + 0.011802741326391697, + 0.018068987876176834, + 0.02119467593729496, + 0.012439110316336155, + 0.01824779622256756, + 0.02496977150440216, + 0.021718228235840797, + 0.003674281295388937, + 0.005608522333204746, + 0.00529844406992197, + 0.004450131207704544, + 0.006947258021682501, + 0.006146403029561043, + 0.00522252032533288, + 0.005335375666618347, + 0.002462932374328375, + 0.0030541412997990847, + 0.00480510201305151, + 0.005083942320197821, + 0.004111400805413723, + 0.004269782453775406, + 0.004915259778499603, + 0.004771087318658829, + 0.009829138405621052, + 0.015315418131649494, + 0.02845444157719612, + 0.019972767680883408, + 0.03552078455686569, + 0.015893978998064995, + 0.022754676640033722, + 0.020772475749254227, + 0.005514151882380247, + 0.005123591516166925, + 0.011524046771228313, + 0.011622526682913303, + 0.0071897245943546295, + 0.015990570187568665, + 0.019260704517364502, + 0.01842326670885086, + 0.0023593867663294077, + 0.003004383994266391, + 0.008444666862487793, + 0.005582976620644331, + 0.004295368678867817, + 0.0037688964512199163, + 0.00524467509239912, + 0.008366295136511326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.22552794218063354 + } + ], + "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/0000084/group_4/index_tensor([4931, 4931])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_tensor([4931, 4931])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f4e02d1d579546cf85bdf07c71e27a24b688bc34 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_tensor([4931, 4931])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle drone shot slowly and steadily pulls back to reveal a vast, boundless enemy formation. Thousands of soldiers, clad in uniform iron armor that gleams with a cold metallic sheen and wearing peaked helmets, stand in impenetrable formations, their spears wielded like a forest of steel rising upward. Their heavy boots, in unison, tread upon the vast, dry, cracked earth. Each step stirs up a swirling cloud of yellow-brown dust, forming a massive wall of sand that threatens to engulf the sky. Illuminated by the glaring midday sun, this torrent of dust takes on a grainy, highly realistic quality, blurring the rear of the army and making it appear endless. Like an unstoppable black tide, the entire formation surges forward with a heavy, relentless rhythm. The countless spearpoints and scratched round iron shields reflect the glaring light, creating a suffocating, epic, oppressive, and apocalyptic atmosphere.", + "seed": 15834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_4931_seed_15834_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": -0.31838574185967444, + "vlm_logprob": -0.31838574185967444, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.31838574185967444, + "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 taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2609595060348511 + ], + "opposite_logprob_tokens": [ + -1.510959506034851 + ], + "logprob_reward": -0.2609595060348511, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2609595060348511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09347515553236008, + 0.047402314841747284, + 0.16282039880752563, + 0.08761125802993774, + 0.05035148188471794, + 0.4235698878765106, + 0.02817663364112377, + 0.03226521611213684, + 0.030254390090703964, + 0.04407321289181709 + ], + "credit_map": [ + 0.01632120832800865, + 0.012111233547329903, + 0.010038181208074093, + 0.009500548243522644, + 0.014531617052853107, + 0.01724068447947502, + 0.006037360057234764, + 0.007694341707974672, + 0.005240553058683872, + 0.0065672085620462894, + 0.004843378439545631, + 0.005031427834182978, + 0.00952343549579382, + 0.007104648742824793, + 0.005973137449473143, + 0.0031185285188257694, + 0.020875131711363792, + 0.021503353491425514, + 0.01713714748620987, + 0.013426648452877998, + 0.033891983330249786, + 0.02625258080661297, + 0.018914684653282166, + 0.010818875394761562, + 0.012740283273160458, + 0.01192533876746893, + 0.009173484519124031, + 0.010330685414373875, + 0.015085835009813309, + 0.01281013060361147, + 0.007895627059042454, + 0.007649880833923817, + 0.009625207632780075, + 0.005760878790169954, + 0.00719138327986002, + 0.00477283401414752, + 0.007053395267575979, + 0.007433227263391018, + 0.004298602696508169, + 0.004215961322188377, + 0.09338168054819107, + 0.06331703066825867, + 0.05020911246538162, + 0.029720157384872437, + 0.08288721740245819, + 0.054316069930791855, + 0.028748493641614914, + 0.02099018357694149, + 0.0034529101103544235, + 0.0035206107422709465, + 0.003268811386078596, + 0.005014749709516764, + 0.0040904078632593155, + 0.0037809195928275585, + 0.0025554231833666563, + 0.0024928043130785227, + 0.0040430715307593346, + 0.002828533761203289, + 0.005018352996557951, + 0.006206536199897528, + 0.0033826841972768307, + 0.004253210499882698, + 0.004308389965444803, + 0.0022244437132030725, + 0.0036294986493885517, + 0.0036074279341846704, + 0.0032873856835067272, + 0.00539585342630744, + 0.0037351218052208424, + 0.0036692647263407707, + 0.004638465121388435, + 0.0022913755383342505, + 0.0032960395328700542, + 0.0052461628802120686, + 0.0043592494912445545, + 0.007240606937557459, + 0.0032801246270537376, + 0.003729785792529583, + 0.013691111467778683, + 0.003230141242966056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.2609595060348511 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373809531331062 + ], + "opposite_logprob_tokens": [ + -3.408738136291504 + ], + "logprob_reward": -0.03373809531331062, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373809531331062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2096678912639618, + 0.04622378572821617, + 0.23978029191493988, + 0.056808553636074066, + 0.03880107030272484, + 0.06447545439004898, + 0.032151028513908386, + 0.014956554397940636, + 0.03790786862373352, + 0.25922760367393494 + ], + "credit_map": [ + 0.030819829553365707, + 0.026050569489598274, + 0.031570740044116974, + 0.02375868521630764, + 0.031525660306215286, + 0.02929951809346676, + 0.01737205497920513, + 0.01927081122994423, + 0.004741624463349581, + 0.007856558077037334, + 0.006642176304012537, + 0.006896099541336298, + 0.00673504127189517, + 0.0043469988740980625, + 0.005128954071551561, + 0.003876328468322754, + 0.029389748349785805, + 0.03185040131211281, + 0.024499772116541862, + 0.015642667189240456, + 0.05047966539859772, + 0.04177122190594673, + 0.03129845857620239, + 0.014848308637738228, + 0.008023967035114765, + 0.008041204884648323, + 0.004054977558553219, + 0.0049983481876552105, + 0.010004150681197643, + 0.013124135322868824, + 0.0031849907245486975, + 0.005376771558076143, + 0.008323265239596367, + 0.0056373355910182, + 0.0031670210883021355, + 0.0028156335465610027, + 0.008039461448788643, + 0.005847254302352667, + 0.0025211696047335863, + 0.002449920866638422, + 0.01430313941091299, + 0.00957342330366373, + 0.008052238263189793, + 0.005004179663956165, + 0.011624650098383427, + 0.008406324312090874, + 0.0040184808894991875, + 0.0034930137917399406, + 0.0036819421220570803, + 0.004281752742826939, + 0.004451078362762928, + 0.006821755785495043, + 0.002739150309935212, + 0.003366137621924281, + 0.0034235597122460604, + 0.003385649062693119, + 0.0020601539872586727, + 0.001412430196069181, + 0.0017629348440095782, + 0.0025671259500086308, + 0.0021825542207807302, + 0.0016304303426295519, + 0.0018702266970649362, + 0.0014706962974742055, + 0.004392857197672129, + 0.004488873761147261, + 0.003064826363697648, + 0.00457449397072196, + 0.004014881327748299, + 0.005645308177918196, + 0.005513460841029882, + 0.0062131634913384914, + 0.011148577556014061, + 0.06176801025867462, + 0.018179157748818398, + 0.015247118659317493, + 0.01784699596464634, + 0.013538649305701256, + 0.10094127058982849, + 0.020557785406708717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.03373809531331062 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233495235443115 + ], + "opposite_logprob_tokens": [ + -0.8983495235443115 + ], + "logprob_reward": -0.5233495235443115, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233495235443115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07950704544782639, + 0.09427914768457413, + 0.257843554019928, + 0.050301071256399155, + 0.07780402153730392, + 0.13517285883426666, + 0.05451253801584244, + 0.018472878262400627, + 0.015495757572352886, + 0.2166111171245575 + ], + "credit_map": [ + 0.01384697388857603, + 0.008889691904187202, + 0.01146064605563879, + 0.00942419096827507, + 0.011364314705133438, + 0.011065577156841755, + 0.0059431795962154865, + 0.007512467913329601, + 0.011399351991713047, + 0.013301393948495388, + 0.009963562712073326, + 0.008501852862536907, + 0.02245010808110237, + 0.012612372636795044, + 0.010464537888765335, + 0.00558595871552825, + 0.035345207899808884, + 0.03515031561255455, + 0.02559291571378708, + 0.01796656847000122, + 0.055406421422958374, + 0.046982135623693466, + 0.02422277256846428, + 0.017177218571305275, + 0.00853402353823185, + 0.006693842355161905, + 0.003820466110482812, + 0.005078084301203489, + 0.007845747284591198, + 0.010258217342197895, + 0.00421003345400095, + 0.0038606571033596992, + 0.01138356700539589, + 0.011980704963207245, + 0.008424673229455948, + 0.004313981160521507, + 0.015768863260746002, + 0.014143765904009342, + 0.006950710900127888, + 0.00483775045722723, + 0.02834424376487732, + 0.02049700915813446, + 0.013504656963050365, + 0.008283264935016632, + 0.028374096378684044, + 0.019584553316235542, + 0.00934402272105217, + 0.007241003215312958, + 0.008806954137980938, + 0.006525349337607622, + 0.008937493897974491, + 0.007106913719326258, + 0.006597368977963924, + 0.007616112008690834, + 0.005124181509017944, + 0.003798162331804633, + 0.0017174130771309137, + 0.0019894358702003956, + 0.0018096689600497484, + 0.0023270940873771906, + 0.002846553921699524, + 0.004599933046847582, + 0.0015952265821397305, + 0.0015875526005402207, + 0.0020405002869665623, + 0.002185468329116702, + 0.0011557863326743245, + 0.001753397285938263, + 0.002587411319836974, + 0.0026019196957349777, + 0.0019086175598204136, + 0.0012626562966033816, + 0.00947589986026287, + 0.05567462742328644, + 0.015741797164082527, + 0.012616184540092945, + 0.013551431708037853, + 0.007393436506390572, + 0.08569720387458801, + 0.01646052673459053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.5233495235443115 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607564091682434 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607564091682434, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607564091682434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09510500729084015, + 0.050349161028862, + 0.172069251537323, + 0.04258081689476967, + 0.06325235217809677, + 0.04221545532345772, + 0.03485560417175293, + 0.05633208900690079, + 0.020757777616381645, + 0.42248252034187317 + ], + "credit_map": [ + 0.01590132713317871, + 0.01101222075521946, + 0.014245844446122646, + 0.009886438958346844, + 0.012685166671872139, + 0.01433615107089281, + 0.006993340793997049, + 0.010044517926871777, + 0.006303684785962105, + 0.0077805472537875175, + 0.005250665824860334, + 0.005746114999055862, + 0.008585506118834019, + 0.007970575243234634, + 0.004423545207828283, + 0.004288524389266968, + 0.022830219939351082, + 0.021918674930930138, + 0.017648063600063324, + 0.012405353598296642, + 0.03665135055780411, + 0.03060714714229107, + 0.017676101997494698, + 0.012332330457866192, + 0.007029104046523571, + 0.006147016305476427, + 0.004154594149440527, + 0.004626117646694183, + 0.0069822450168430805, + 0.006131614558398724, + 0.003141334280371666, + 0.004368787165731192, + 0.010195357725024223, + 0.008001528680324554, + 0.005066251847893, + 0.004291143734008074, + 0.011589069850742817, + 0.01384028885513544, + 0.004731622990220785, + 0.005537089426070452, + 0.008589653298258781, + 0.006423603277653456, + 0.004841494839638472, + 0.0028221739921718836, + 0.008769205771386623, + 0.005866662133485079, + 0.0028676320798695087, + 0.002035029698163271, + 0.008677590638399124, + 0.004647605121135712, + 0.005362067837268114, + 0.003106597810983658, + 0.004458539187908173, + 0.0035651966463774443, + 0.002306184032931924, + 0.0027318191714584827, + 0.007736892905086279, + 0.0060698725283145905, + 0.0069122607819736, + 0.010289190337061882, + 0.008346115238964558, + 0.006501965690404177, + 0.006385080516338348, + 0.004090713337063789, + 0.0022721330169588327, + 0.0023915739730000496, + 0.002539156936109066, + 0.004219437949359417, + 0.0018802803242579103, + 0.0022450927644968033, + 0.002511783968657255, + 0.002698317402973771, + 0.017099104821681976, + 0.11194906383752823, + 0.0292140394449234, + 0.024420756846666336, + 0.027075204998254776, + 0.012270396575331688, + 0.16895127296447754, + 0.03150266408920288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.01607564091682434 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578059434890747 + ], + "opposite_logprob_tokens": [ + -0.6328059434890747 + ], + "logprob_reward": -0.7578059434890747, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578059434890747, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10918933153152466, + 0.0958670973777771, + 0.12189831584692001, + 0.060424335300922394, + 0.1530100554227829, + 0.22266562283039093, + 0.13577626645565033, + 0.036815088242292404, + 0.030627621337771416, + 0.033726245164871216 + ], + "credit_map": [ + 0.013020371086895466, + 0.013431928120553493, + 0.014645867981016636, + 0.013235382735729218, + 0.019989723339676857, + 0.014028741046786308, + 0.010737725533545017, + 0.010099594481289387, + 0.012806421145796776, + 0.012206369079649448, + 0.01035686768591404, + 0.010466909036040306, + 0.01498856209218502, + 0.017169583588838577, + 0.009134199470281601, + 0.008738178759813309, + 0.014390434138476849, + 0.01630779355764389, + 0.010577598586678505, + 0.012507649138569832, + 0.02532155252993107, + 0.019522899761795998, + 0.014216504991054535, + 0.00905388779938221, + 0.006939314771443605, + 0.007823050953447819, + 0.004850375931710005, + 0.003549333894625306, + 0.006965855602174997, + 0.014461629092693329, + 0.0064924415200948715, + 0.009342333301901817, + 0.02843671478331089, + 0.020749429240822792, + 0.014642264693975449, + 0.010263800621032715, + 0.03115963749587536, + 0.02583410032093525, + 0.01132322195917368, + 0.01060088723897934, + 0.0460384264588356, + 0.03386898711323738, + 0.022255221381783485, + 0.014368592761456966, + 0.046544790267944336, + 0.02987324260175228, + 0.017382917925715446, + 0.012333443388342857, + 0.02860438823699951, + 0.019291365519165993, + 0.018954288214445114, + 0.009516991674900055, + 0.019430598244071007, + 0.024407781660556793, + 0.007874337956309319, + 0.00769650237634778, + 0.004117749631404877, + 0.0038839238695800304, + 0.00530823552981019, + 0.009554841555655003, + 0.0036946407053619623, + 0.003826557192951441, + 0.004058314021676779, + 0.002370822476223111, + 0.002763037569820881, + 0.0034731596242636442, + 0.0025979969650506973, + 0.005060915369540453, + 0.0038387428503483534, + 0.006536644883453846, + 0.0034391412045806646, + 0.002917984500527382, + 0.003544711507856846, + 0.007509106770157814, + 0.0032757504377514124, + 0.0033724154345691204, + 0.004705918487161398, + 0.00433529494330287, + 0.0042332857847213745, + 0.0027497627306729555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.7578059434890747 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle drone shot slowly and steadily pulls back to reveal a vast, boundless enemy formation. Thousands of soldiers, clad in uniform iron armor that gleams with a cold metallic sheen and wearing peaked helmets, stand in impenetrable formations, their spears wielded like a forest of steel rising upward. Their heavy boots, in unison, tread upon the vast, dry, cracked earth. Each step stirs up a swirling cloud of yellow-brown dust, forming a massive wall of sand that threatens to engulf the sky. Illuminated by the glaring midday sun, this torrent of dust takes on a grainy, highly realistic quality, blurring the rear of the army and making it appear endless. Like an unstoppable black tide, the entire formation surges forward with a heavy, relentless rhythm. The countless spearpoints and scratched round iron shields reflect the glaring light, creating a suffocating, epic, oppressive, and apocalyptic atmosphere.", + "seed": 15835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_4931_seed_15835_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": -0.35866222307085993, + "vlm_logprob": -0.35866222307085993, + "vlm_yes_no_margin": 1.5249999523162843, + "vlm_token_credit": -0.35866222307085993, + "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 taken from a high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8471368551254272 + ], + "opposite_logprob_tokens": [ + -0.5971368551254272 + ], + "logprob_reward": -0.8471368551254272, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8471368551254272, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.196626678109169, + 0.0922398492693901, + 0.06773108243942261, + 0.16881734132766724, + 0.08163167536258698, + 0.034801311790943146, + 0.04228504002094269, + 0.042622972279787064, + 0.09324970841407776, + 0.17999432981014252 + ], + "credit_map": [ + 0.018166877329349518, + 0.06169316545128822, + 0.009094420820474625, + 0.009862121194601059, + 0.04688717797398567, + 0.02631540596485138, + 0.015254128724336624, + 0.009353379718959332, + 0.010158096440136433, + 0.01081025879830122, + 0.012972773984074593, + 0.010084475390613079, + 0.013044859282672405, + 0.015288673341274261, + 0.01013906765729189, + 0.009741649031639099, + 0.008356964215636253, + 0.006302418187260628, + 0.006715044379234314, + 0.007592245936393738, + 0.012436194345355034, + 0.011859911493957043, + 0.007834752090275288, + 0.006633555982261896, + 0.030435901135206223, + 0.01133066788315773, + 0.020266663283109665, + 0.010565640404820442, + 0.020763568580150604, + 0.02092478796839714, + 0.01987113244831562, + 0.03465898707509041, + 0.007981712929904461, + 0.007078798953443766, + 0.013276097364723682, + 0.00809470098465681, + 0.010843262076377869, + 0.017782671377062798, + 0.008053820580244064, + 0.008520613424479961, + 0.005437880754470825, + 0.0036380405072122812, + 0.004317346960306168, + 0.0031630212906748056, + 0.005017520394176245, + 0.005171351134777069, + 0.004354520700871944, + 0.003701629117131233, + 0.005166460294276476, + 0.0042285495437681675, + 0.006930773612111807, + 0.005014790687710047, + 0.00732042733579874, + 0.00714915432035923, + 0.003393427701666951, + 0.003081453265622258, + 0.0047099776566028595, + 0.00556504912674427, + 0.003908458165824413, + 0.0046619693748652935, + 0.005091393366456032, + 0.010333307087421417, + 0.004579675383865833, + 0.003773145144805312, + 0.011718617752194405, + 0.007225419860333204, + 0.008284458890557289, + 0.009322856552898884, + 0.010134345851838589, + 0.02071642316877842, + 0.012816434726119041, + 0.013031148351728916, + 0.014101900160312653, + 0.01876913197338581, + 0.015343568287789822, + 0.02867119014263153, + 0.016652364283800125, + 0.02260539121925831, + 0.024780770763754845, + 0.0390700101852417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.8471368551254272 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07009705901145935 + ], + "opposite_logprob_tokens": [ + -2.695096969604492 + ], + "logprob_reward": -0.07009705901145935, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07009705901145935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22695958614349365, + 0.05937475338578224, + 0.09517833590507507, + 0.2188417613506317, + 0.052999451756477356, + 0.06708496063947678, + 0.05482259392738342, + 0.05393335223197937, + 0.10765180736780167, + 0.06315339356660843 + ], + "credit_map": [ + 0.026688333600759506, + 0.046145278960466385, + 0.018613820895552635, + 0.019951455295085907, + 0.04036858305335045, + 0.024902109056711197, + 0.023984601721167564, + 0.026305392384529114, + 0.0058502862229943275, + 0.0055292039178311825, + 0.004945663269609213, + 0.0036595899146050215, + 0.014695709571242332, + 0.015638219192624092, + 0.004923662636429071, + 0.004132416564971209, + 0.009648923762142658, + 0.010049338452517986, + 0.008502681739628315, + 0.008915767073631287, + 0.017618348821997643, + 0.01937474124133587, + 0.013286055997014046, + 0.007782480213791132, + 0.04293728992342949, + 0.016119088977575302, + 0.02451889030635357, + 0.0175953209400177, + 0.02741382271051407, + 0.026421668007969856, + 0.02506062760949135, + 0.03877505287528038, + 0.0049424804747104645, + 0.005969592835754156, + 0.0066573419608175755, + 0.009116311557590961, + 0.006887729745358229, + 0.009056161157786846, + 0.004776919726282358, + 0.0055929142981767654, + 0.005056147463619709, + 0.00604122132062912, + 0.005739338230341673, + 0.005048890598118305, + 0.019425591453909874, + 0.016079751774668694, + 0.005175097845494747, + 0.004518923349678516, + 0.0063644954934716225, + 0.008404677733778954, + 0.0046945721842348576, + 0.006352831143885851, + 0.009380538947880268, + 0.010407212190330029, + 0.005507422145456076, + 0.003710847580805421, + 0.006076517980545759, + 0.006846373435109854, + 0.004439572338014841, + 0.0063992030918598175, + 0.006859899498522282, + 0.011612026020884514, + 0.006618613842874765, + 0.005081143230199814, + 0.020824017003178596, + 0.010084227658808231, + 0.011274021118879318, + 0.012552576139569283, + 0.011437719687819481, + 0.014487044885754585, + 0.013584326021373272, + 0.013407875783741474, + 0.005832813214510679, + 0.008758766576647758, + 0.006489485502243042, + 0.011228724382817745, + 0.00889368075877428, + 0.007401564158499241, + 0.008341661654412746, + 0.006206695456057787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.07009705901145935 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42895594239234924 + ], + "opposite_logprob_tokens": [ + -1.0539559125900269 + ], + "logprob_reward": -0.42895594239234924, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42895594239234924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5537726283073425, + 0.024046804755926132, + 0.07213222980499268, + 0.04603471979498863, + 0.016923459246754646, + 0.04102061316370964, + 0.03451160341501236, + 0.039889298379421234, + 0.05381203070282936, + 0.11785667389631271 + ], + "credit_map": [ + 0.05321906507015228, + 0.1757005900144577, + 0.017249317839741707, + 0.023495597764849663, + 0.059497471898794174, + 0.16011221706867218, + 0.02935095876455307, + 0.035147424787282944, + 0.003175092861056328, + 0.003513369942083955, + 0.0025046514347195625, + 0.0022315082605928183, + 0.004663958214223385, + 0.003914752975106239, + 0.0019113562302663922, + 0.0021321133244782686, + 0.009336380288004875, + 0.0057791173458099365, + 0.005998639855533838, + 0.004630906507372856, + 0.00973063986748457, + 0.02155628427863121, + 0.006489274092018604, + 0.008610992692410946, + 0.004732378292828798, + 0.005529072601348162, + 0.004580534063279629, + 0.006455575581640005, + 0.004763483069837093, + 0.008420456200838089, + 0.0068674953654408455, + 0.004685722291469574, + 0.001765757566317916, + 0.0015671307919546962, + 0.0023141249548643827, + 0.0016674724174663424, + 0.0024216342717409134, + 0.0025336958933621645, + 0.0026914949994534254, + 0.0019621471874415874, + 0.0034835096448659897, + 0.0036116086412221193, + 0.006658820901066065, + 0.004313678480684757, + 0.007057917304337025, + 0.004369854927062988, + 0.008318266831338406, + 0.003206953639164567, + 0.00366977509111166, + 0.0031892431434243917, + 0.002296747639775276, + 0.0027598775923252106, + 0.011364547535777092, + 0.005968304350972176, + 0.0026412992738187313, + 0.002621806226670742, + 0.0032154866494238377, + 0.0037670177407562733, + 0.002912205643951893, + 0.004728879313915968, + 0.006033840123564005, + 0.006637285929173231, + 0.008886474184691906, + 0.003708108328282833, + 0.009714983403682709, + 0.006144693586975336, + 0.003977485932409763, + 0.005847695283591747, + 0.00548361474648118, + 0.007768966723233461, + 0.007697927765548229, + 0.007176664192229509, + 0.010485243052244186, + 0.012720152735710144, + 0.008708856999874115, + 0.014093950390815735, + 0.015072613023221493, + 0.015207063406705856, + 0.01560711394995451, + 0.025961685925722122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.42895594239234924 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201585859060287 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201585859060287, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018201585859060287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39924129843711853, + 0.028240405023097992, + 0.060467369854450226, + 0.07796888798475266, + 0.026535483077168465, + 0.17128832638263702, + 0.05274154990911484, + 0.11792699992656708, + 0.03242385759949684, + 0.033165816217660904 + ], + "credit_map": [ + 0.03283866122364998, + 0.10082700848579407, + 0.012080192565917969, + 0.016740506514906883, + 0.035696398466825485, + 0.15241926908493042, + 0.024184536188840866, + 0.024454738944768906, + 0.0023768306709825993, + 0.002390369074419141, + 0.0028922560159116983, + 0.0020120383705943823, + 0.004769056104123592, + 0.006965539418160915, + 0.003015438560396433, + 0.003818874480202794, + 0.009575282223522663, + 0.003998931962996721, + 0.005595133174210787, + 0.004111144691705704, + 0.009546413086354733, + 0.010628442279994488, + 0.005343622528016567, + 0.011668399907648563, + 0.01576986536383629, + 0.007065851241350174, + 0.0072051649913191795, + 0.007033615373075008, + 0.00928674265742302, + 0.00950725469738245, + 0.009960219264030457, + 0.012140175327658653, + 0.003208801383152604, + 0.0023869446013122797, + 0.004355043172836304, + 0.002800658345222473, + 0.0038045577239245176, + 0.0034745358861982822, + 0.003919590264558792, + 0.00258535030297935, + 0.019856980070471764, + 0.013033300638198853, + 0.01780928671360016, + 0.009724360890686512, + 0.03812289237976074, + 0.03581063076853752, + 0.016705462709069252, + 0.020225413143634796, + 0.004891176242381334, + 0.005354582332074642, + 0.008560880087316036, + 0.005114009603857994, + 0.008818594738841057, + 0.011683065444231033, + 0.0038954922929406166, + 0.004423750098794699, + 0.006422846578061581, + 0.008098985999822617, + 0.004792940802872181, + 0.015082055702805519, + 0.030918661504983902, + 0.010372065007686615, + 0.03520159795880318, + 0.007037849631160498, + 0.00427643395960331, + 0.0027774323243647814, + 0.0034442527685314417, + 0.003499941900372505, + 0.003730306401848793, + 0.006203440949320793, + 0.004280409775674343, + 0.004211637657135725, + 0.0033767055720090866, + 0.005941763985902071, + 0.0038817026652395725, + 0.004985354375094175, + 0.004854880273342133, + 0.0030090808868408203, + 0.003674101084470749, + 0.003442229935899377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.018201585859060287 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289196729660034 + ], + "opposite_logprob_tokens": [ + -1.0539196729660034 + ], + "logprob_reward": -0.4289196729660034, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289196729660034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3716620206832886, + 0.02930820919573307, + 0.035791970789432526, + 0.12650786340236664, + 0.02618798427283764, + 0.03616290166974068, + 0.039919860661029816, + 0.14318430423736572, + 0.07539801299571991, + 0.11587684601545334 + ], + "credit_map": [ + 0.04199643433094025, + 0.08226948976516724, + 0.018277131021022797, + 0.025159912183880806, + 0.03655746206641197, + 0.11128924041986465, + 0.025572484359145164, + 0.03053990751504898, + 0.00314009515568614, + 0.004178488627076149, + 0.0024654720909893513, + 0.0020779017359018326, + 0.005176321137696505, + 0.0069763753563165665, + 0.003286304185166955, + 0.002007250441238284, + 0.003649605205282569, + 0.0035189841873943806, + 0.003171235788613558, + 0.0026601948775351048, + 0.006117179058492184, + 0.010540639981627464, + 0.003653559135273099, + 0.0024805779103189707, + 0.02106546238064766, + 0.011172059923410416, + 0.014944287948310375, + 0.016956016421318054, + 0.01847033202648163, + 0.020686864852905273, + 0.013501918874680996, + 0.009710936807096004, + 0.002040094928815961, + 0.0026458122301846743, + 0.003712001256644726, + 0.0023786642123013735, + 0.004313732963055372, + 0.004147494211792946, + 0.004126403946429491, + 0.0028237830847501755, + 0.0038161934353411198, + 0.004167378414422274, + 0.004501137882471085, + 0.004037506878376007, + 0.006375596392899752, + 0.006117571145296097, + 0.0033126489724963903, + 0.0038348701782524586, + 0.003679305547848344, + 0.0038217552937567234, + 0.00577946612611413, + 0.003963092342019081, + 0.007334594614803791, + 0.009057199582457542, + 0.0037101656198501587, + 0.0025742840953171253, + 0.010421162471175194, + 0.01218041218817234, + 0.013361326418817043, + 0.017135029658675194, + 0.026324568316340446, + 0.018287358805537224, + 0.031952597200870514, + 0.013521861284971237, + 0.008952312171459198, + 0.007138968911021948, + 0.010655907914042473, + 0.009378865361213684, + 0.009409346617758274, + 0.009684648364782333, + 0.010240987874567509, + 0.009936985559761524, + 0.011580605991184711, + 0.009036501869559288, + 0.011170036159455776, + 0.011343077756464481, + 0.0148519491776824, + 0.016829557716846466, + 0.015144015662372112, + 0.025921106338500977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.4289196729660034 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot slowly and steadily pulls back to reveal a vast, boundless enemy formation. Thousands of soldiers, clad in uniform iron armor that gleams with a cold metallic sheen and wearing peaked helmets, stand in impenetrable formations, their spears wielded like a forest of steel rising upward. Their heavy boots, in unison, tread upon the vast, dry, cracked earth. Each step stirs up a swirling cloud of yellow-brown dust, forming a massive wall of sand that threatens to engulf the sky. Illuminated by the glaring midday sun, this torrent of dust takes on a grainy, highly realistic quality, blurring the rear of the army and making it appear endless. Like an unstoppable black tide, the entire formation surges forward with a heavy, relentless rhythm. The countless spearpoints and scratched round iron shields reflect the glaring light, creating a suffocating, epic, oppressive, and apocalyptic atmosphere.", + "seed": 15836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_4931_seed_15836_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": -0.2824943091720343, + "vlm_logprob": -0.2824943091720343, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.2824943091720343, + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6488675475120544 + ], + "opposite_logprob_tokens": [ + -0.7738675475120544 + ], + "logprob_reward": -0.6488675475120544, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6488675475120544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19495819509029388, + 0.14376698434352875, + 0.06704860925674438, + 0.08051140606403351, + 0.0970621258020401, + 0.18917317688465118, + 0.0461764894425869, + 0.09447585046291351, + 0.03660045936703682, + 0.05022672563791275 + ], + "credit_map": [ + 0.029119903221726418, + 0.023204192519187927, + 0.0214382354170084, + 0.027712784707546234, + 0.01914355717599392, + 0.03506717458367348, + 0.018935224041342735, + 0.020337112247943878, + 0.010845191776752472, + 0.027145054191350937, + 0.01559702679514885, + 0.017485840246081352, + 0.015237096697092056, + 0.015508999116718769, + 0.027389662340283394, + 0.014558112248778343, + 0.00795760564506054, + 0.006855528336018324, + 0.009364486671984196, + 0.00897999107837677, + 0.00831681303679943, + 0.007954888045787811, + 0.009101586416363716, + 0.008517714217305183, + 0.009106185287237167, + 0.012404272332787514, + 0.008868773467838764, + 0.011702721938490868, + 0.00981463585048914, + 0.009714166633784771, + 0.010748328641057014, + 0.008152322843670845, + 0.013328290544450283, + 0.012137480080127716, + 0.007846605964004993, + 0.008645297028124332, + 0.01943095214664936, + 0.017903480678796768, + 0.009217177517712116, + 0.008552837185561657, + 0.019725294783711433, + 0.023059139028191566, + 0.020117972046136856, + 0.017444532364606857, + 0.05022053420543671, + 0.021053658798336983, + 0.018194301053881645, + 0.019357744604349136, + 0.005425743293017149, + 0.005457036662846804, + 0.004078011494129896, + 0.004522150848060846, + 0.00749854976311326, + 0.0069379075430333614, + 0.006113776471465826, + 0.0061433156952261925, + 0.016188159584999084, + 0.010325520299375057, + 0.011262288317084312, + 0.01082039438188076, + 0.012979576364159584, + 0.009608032181859016, + 0.012450497597455978, + 0.010841384530067444, + 0.0032713976688683033, + 0.004738130606710911, + 0.004268723074346781, + 0.005335352849215269, + 0.0045337677001953125, + 0.0030478297267109156, + 0.00583708006888628, + 0.005568178370594978, + 0.004837240092456341, + 0.004617629107087851, + 0.007681573741137981, + 0.010156813077628613, + 0.0038836209569126368, + 0.004869441036134958, + 0.006750671658664942, + 0.007429736666381359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.6488675475120544 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11301614344120026 + ], + "opposite_logprob_tokens": [ + -2.238016128540039 + ], + "logprob_reward": -0.11301614344120026, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11301614344120026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1634281575679779, + 0.15847083926200867, + 0.05116276443004608, + 0.04464587941765785, + 0.05278603360056877, + 0.12175095826387405, + 0.18780069053173065, + 0.07949390262365341, + 0.05685766786336899, + 0.08360309898853302 + ], + "credit_map": [ + 0.01915052905678749, + 0.022937487810850143, + 0.02579217404127121, + 0.019178470596671104, + 0.015245258808135986, + 0.027445552870631218, + 0.01695227436721325, + 0.016726404428482056, + 0.01443439070135355, + 0.029762325808405876, + 0.015693748369812965, + 0.016919054090976715, + 0.01781986467540264, + 0.018052328377962112, + 0.03196711465716362, + 0.013822004199028015, + 0.004733382258564234, + 0.005345612298697233, + 0.005531758069992065, + 0.008252467028796673, + 0.0056845867075026035, + 0.004713132046163082, + 0.009529252536594868, + 0.007372566033154726, + 0.005598908755928278, + 0.00662263622507453, + 0.004432256333529949, + 0.0053037903271615505, + 0.00658315047621727, + 0.006219388451427221, + 0.004735201597213745, + 0.005150544457137585, + 0.0059927101247012615, + 0.005765705835074186, + 0.0051302737556397915, + 0.005723575130105019, + 0.007783437613397837, + 0.005862989462912083, + 0.007553430274128914, + 0.008973910473287106, + 0.01208783220499754, + 0.0150647833943367, + 0.013994117267429829, + 0.009158728644251823, + 0.03935832530260086, + 0.013089503161609173, + 0.011057383380830288, + 0.007940277457237244, + 0.016823869198560715, + 0.029414262622594833, + 0.011665873229503632, + 0.013623367995023727, + 0.02675878256559372, + 0.02393730729818344, + 0.02438472770154476, + 0.04119247943162918, + 0.011075451970100403, + 0.009955352172255516, + 0.009516110643744469, + 0.010407350026071072, + 0.010123475454747677, + 0.006752734538167715, + 0.011227822862565517, + 0.010435599833726883, + 0.005642001982778311, + 0.005259544588625431, + 0.007908930070698261, + 0.0094785550609231, + 0.005409353878349066, + 0.005100633017718792, + 0.009873349219560623, + 0.00818529911339283, + 0.00974898599088192, + 0.008312273770570755, + 0.008838439360260963, + 0.011701954528689384, + 0.007166487630456686, + 0.008739079348742962, + 0.0135775962844491, + 0.015518276952207088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.11301614344120026 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486623764038086 + ], + "opposite_logprob_tokens": [ + -1.2236623764038086 + ], + "logprob_reward": -0.3486623764038086, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486623764038086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15410566329956055, + 0.05631260201334953, + 0.09668854624032974, + 0.05690959841012955, + 0.04495582729578018, + 0.02924947626888752, + 0.09486287832260132, + 0.07666188478469849, + 0.31971070170402527, + 0.07054286450147629 + ], + "credit_map": [ + 0.017941901460289955, + 0.025372158735990524, + 0.01618160679936409, + 0.01864400878548622, + 0.022952990606427193, + 0.021572211757302284, + 0.015266361646354198, + 0.016174418851733208, + 0.003909323364496231, + 0.012121659703552723, + 0.006985239218920469, + 0.007334526628255844, + 0.006299891043454409, + 0.004462100565433502, + 0.010234255343675613, + 0.004965601023286581, + 0.011631492525339127, + 0.01276673749089241, + 0.011202954687178135, + 0.01108099427074194, + 0.016082603484392166, + 0.01060893852263689, + 0.012784353457391262, + 0.01053046714514494, + 0.006170167122036219, + 0.007621023803949356, + 0.006516629364341497, + 0.008760245516896248, + 0.006473488174378872, + 0.008712287060916424, + 0.00720239570364356, + 0.005453362129628658, + 0.005151218734681606, + 0.006044268608093262, + 0.00448561180382967, + 0.005352832842618227, + 0.005559672135859728, + 0.004767062608152628, + 0.005904071498662233, + 0.007691090926527977, + 0.0025270043406635523, + 0.005028845276683569, + 0.003189428010955453, + 0.0027437438257038593, + 0.0073022968135774136, + 0.0033327871933579445, + 0.0024855525698512793, + 0.002639820333570242, + 0.009451077319681644, + 0.010993530973792076, + 0.00860615074634552, + 0.006802668329328299, + 0.016998330131173134, + 0.015188975259661674, + 0.009821271523833275, + 0.017000870779156685, + 0.010822569951415062, + 0.007341369986534119, + 0.00759825250133872, + 0.007248294539749622, + 0.014835421927273273, + 0.012206313200294971, + 0.008242196403443813, + 0.008367467671632767, + 0.027494113892316818, + 0.0439472459256649, + 0.022554881870746613, + 0.031706083565950394, + 0.02858962118625641, + 0.0212063267827034, + 0.07904659956693649, + 0.06516585499048233, + 0.009053201414644718, + 0.00596494460478425, + 0.0068723601289093494, + 0.008263331837952137, + 0.005341336131095886, + 0.006321432534605265, + 0.013583682477474213, + 0.015142573975026608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.3486623764038086 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596204325556755 + ], + "opposite_logprob_tokens": [ + -3.8955962657928467 + ], + "logprob_reward": -0.020596204325556755, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596204325556755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.105955109000206, + 0.06777381151914597, + 0.05176666006445885, + 0.3076942563056946, + 0.04006638750433922, + 0.019152166321873665, + 0.06412053853273392, + 0.02956245467066765, + 0.10480333119630814, + 0.20910534262657166 + ], + "credit_map": [ + 0.010734684765338898, + 0.017800461500883102, + 0.012560239061713219, + 0.013796004466712475, + 0.012159256264567375, + 0.014094943180680275, + 0.013352376408874989, + 0.011457138694822788, + 0.005330437328666449, + 0.015000165440142155, + 0.006480276118963957, + 0.007885916158556938, + 0.0077753993682563305, + 0.00796475075185299, + 0.011264849454164505, + 0.006072011776268482, + 0.006884708534926176, + 0.0061289346776902676, + 0.005438585765659809, + 0.005948151461780071, + 0.007819042541086674, + 0.005872716661542654, + 0.007452333811670542, + 0.006222182884812355, + 0.039277758449316025, + 0.04502641409635544, + 0.035454899072647095, + 0.04598004370927811, + 0.04171101003885269, + 0.03016582690179348, + 0.040546197444200516, + 0.029532087966799736, + 0.005417616106569767, + 0.005928965751081705, + 0.004307233262807131, + 0.007419927511364222, + 0.004650406539440155, + 0.004546868149191141, + 0.004602961707860231, + 0.003192401956766844, + 0.0017478592926636338, + 0.002264657523483038, + 0.002155766123905778, + 0.002081669168546796, + 0.0044851284474134445, + 0.002028492745012045, + 0.002313934499397874, + 0.002074658405035734, + 0.00664858240634203, + 0.007285001687705517, + 0.005400128662586212, + 0.0047331009991467, + 0.01343887485563755, + 0.01054852269589901, + 0.006588933989405632, + 0.009477387182414532, + 0.0029521977994590998, + 0.004460069816559553, + 0.003181541571393609, + 0.0036052430514246225, + 0.003993692807853222, + 0.0024810044560581446, + 0.004311443772166967, + 0.004577259998768568, + 0.010219743475317955, + 0.009381085634231567, + 0.007076984737068415, + 0.011662700213491917, + 0.010191000998020172, + 0.01232204306870699, + 0.02173428051173687, + 0.022215479984879494, + 0.017433878034353256, + 0.01247249636799097, + 0.01751040481030941, + 0.022359753027558327, + 0.015214738436043262, + 0.014448828063905239, + 0.05074475333094597, + 0.05892046168446541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.020596204325556755 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132927417755127 + ], + "opposite_logprob_tokens": [ + -1.4063292741775513 + ], + "logprob_reward": -0.28132927417755127, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132927417755127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15002822875976562, + 0.0640801191329956, + 0.07118251919746399, + 0.06471934169530869, + 0.12537142634391785, + 0.03630649298429489, + 0.07124833017587662, + 0.10389196872711182, + 0.2141307145357132, + 0.09904086589813232 + ], + "credit_map": [ + 0.017188629135489464, + 0.023996401578187943, + 0.022307097911834717, + 0.02097533643245697, + 0.014211959205567837, + 0.01871437579393387, + 0.016821246594190598, + 0.01581317000091076, + 0.005278675351291895, + 0.010792807675898075, + 0.007906917482614517, + 0.008233520202338696, + 0.007855324074625969, + 0.0060843052342534065, + 0.011331962421536446, + 0.006596605759114027, + 0.013488511554896832, + 0.009462560527026653, + 0.008180509321391582, + 0.007770425174385309, + 0.009008413180708885, + 0.009153290651738644, + 0.006996951997280121, + 0.007121860049664974, + 0.006749831140041351, + 0.009749108925461769, + 0.008204974234104156, + 0.008910742588341236, + 0.007887670770287514, + 0.007932575419545174, + 0.0074271573685109615, + 0.00785728171467781, + 0.013018622063100338, + 0.018855402246117592, + 0.013021530583500862, + 0.014009296894073486, + 0.024489520117640495, + 0.018772760406136513, + 0.011580548249185085, + 0.011623742058873177, + 0.0034210237208753824, + 0.0038631476927548647, + 0.00442558852955699, + 0.004543505143374205, + 0.007056172471493483, + 0.004797123372554779, + 0.003958316519856453, + 0.004241615999490023, + 0.0075967940501868725, + 0.007048953790217638, + 0.006162996869534254, + 0.005214367527514696, + 0.011909417808055878, + 0.011644335463643074, + 0.007993596605956554, + 0.013677870854735374, + 0.01391662284731865, + 0.013842344284057617, + 0.011809755116701126, + 0.016372976824641228, + 0.01151348277926445, + 0.009353348053991795, + 0.012587801553308964, + 0.01449564378708601, + 0.02059762552380562, + 0.018713323399424553, + 0.018493402749300003, + 0.0333411805331707, + 0.023904651403427124, + 0.0355614572763443, + 0.02883012592792511, + 0.03468892723321915, + 0.012599200941622257, + 0.009999720379710197, + 0.010406623594462872, + 0.015199071727693081, + 0.00917978398501873, + 0.00861193798482418, + 0.011585663072764874, + 0.021458858624100685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.28132927417755127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot slowly and steadily pulls back to reveal a vast, boundless enemy formation. Thousands of soldiers, clad in uniform iron armor that gleams with a cold metallic sheen and wearing peaked helmets, stand in impenetrable formations, their spears wielded like a forest of steel rising upward. Their heavy boots, in unison, tread upon the vast, dry, cracked earth. Each step stirs up a swirling cloud of yellow-brown dust, forming a massive wall of sand that threatens to engulf the sky. Illuminated by the glaring midday sun, this torrent of dust takes on a grainy, highly realistic quality, blurring the rear of the army and making it appear endless. Like an unstoppable black tide, the entire formation surges forward with a heavy, relentless rhythm. The countless spearpoints and scratched round iron shields reflect the glaring light, creating a suffocating, epic, oppressive, and apocalyptic atmosphere.", + "seed": 15837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_4/index_4931_seed_15837_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": -0.2830863788723946, + "vlm_logprob": -0.2830863788723946, + "vlm_yes_no_margin": 1.500000011920929, + "vlm_token_credit": -0.2830863788723946, + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29130101203918457 + ], + "opposite_logprob_tokens": [ + -1.4163010120391846 + ], + "logprob_reward": -0.29130101203918457, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29130101203918457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2198038101196289, + 0.09555428475141525, + 0.10649946331977844, + 0.044296614825725555, + 0.057632990181446075, + 0.05687017738819122, + 0.09262211620807648, + 0.042015720158815384, + 0.24925386905670166, + 0.03545093163847923 + ], + "credit_map": [ + 0.03612871095538139, + 0.03448121249675751, + 0.016961587592959404, + 0.0210170466452837, + 0.03197222948074341, + 0.032952260226011276, + 0.022427964955568314, + 0.023862790316343307, + 0.011446364223957062, + 0.009112076833844185, + 0.012292037717998028, + 0.009673121385276318, + 0.01950056664645672, + 0.012411669827997684, + 0.011862209066748619, + 0.009256234392523766, + 0.008741891011595726, + 0.01657467894256115, + 0.009112966246902943, + 0.009378858841955662, + 0.02890991047024727, + 0.011196164414286613, + 0.011289168149232864, + 0.011295827105641365, + 0.007809129543602467, + 0.00927826389670372, + 0.0051515838131308556, + 0.0029535682406276464, + 0.005502202082425356, + 0.006585989613085985, + 0.0034376897383481264, + 0.003578188130632043, + 0.007840187288820744, + 0.008472573012113571, + 0.006688382942229509, + 0.005913400091230869, + 0.010389245115220547, + 0.009743327274918556, + 0.0039862580597400665, + 0.004599617328494787, + 0.01094147376716137, + 0.008608578704297543, + 0.007422152906656265, + 0.006162253674119711, + 0.008697873912751675, + 0.0069126226007938385, + 0.0041513461619615555, + 0.003973873797804117, + 0.011296924203634262, + 0.011986401863396168, + 0.009918143972754478, + 0.012164845131337643, + 0.013802031986415386, + 0.01098397746682167, + 0.014729304239153862, + 0.007740488275885582, + 0.0031875527929514647, + 0.011733624152839184, + 0.0026983690913766623, + 0.003048982936888933, + 0.011549580842256546, + 0.004540045280009508, + 0.0025568052660673857, + 0.002700760494917631, + 0.03588608652353287, + 0.06120110675692558, + 0.02669254131615162, + 0.014813660643994808, + 0.05498241260647774, + 0.02359778620302677, + 0.013634505681693554, + 0.01844579167664051, + 0.004263726063072681, + 0.005333854816854, + 0.003258757060393691, + 0.0028742440044879913, + 0.008166578598320484, + 0.006561071611940861, + 0.0024845837615430355, + 0.002508114092051983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.29130101203918457 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12714149057865143 + ], + "opposite_logprob_tokens": [ + -2.1271414756774902 + ], + "logprob_reward": -0.12714149057865143, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12714149057865143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14582158625125885, + 0.08394760638475418, + 0.1012062355875969, + 0.21148887276649475, + 0.056521639227867126, + 0.05535963177680969, + 0.14169934391975403, + 0.06149943172931671, + 0.1175004243850708, + 0.02495524473488331 + ], + "credit_map": [ + 0.017133304849267006, + 0.016274919733405113, + 0.014918752945959568, + 0.015609320253133774, + 0.02664872817695141, + 0.022004593163728714, + 0.017242630943655968, + 0.01598934456706047, + 0.012794478796422482, + 0.008521352894604206, + 0.009163076989352703, + 0.008107727393507957, + 0.01546548306941986, + 0.01255730725824833, + 0.007749284617602825, + 0.009588895365595818, + 0.012239716947078705, + 0.013147706165909767, + 0.00820804201066494, + 0.012993386015295982, + 0.022227654233574867, + 0.01249938365072012, + 0.011063981801271439, + 0.008826370351016521, + 0.04730565845966339, + 0.0331309512257576, + 0.017642350867390633, + 0.024578601121902466, + 0.026361411437392235, + 0.022966276854276657, + 0.02237294614315033, + 0.017130689695477486, + 0.011670485138893127, + 0.01003224216401577, + 0.005149805918335915, + 0.005501806735992432, + 0.007577788084745407, + 0.007517015095800161, + 0.0044901552610099316, + 0.004582344088703394, + 0.008258737623691559, + 0.007620405405759811, + 0.005667864810675383, + 0.0056150504387915134, + 0.007050983142107725, + 0.011552217416465282, + 0.004301463253796101, + 0.005292912945151329, + 0.017535123974084854, + 0.01878943294286728, + 0.014745200984179974, + 0.01743892952799797, + 0.022737370803952217, + 0.0168521199375391, + 0.020897334441542625, + 0.012703834101557732, + 0.003890456398949027, + 0.019649332389235497, + 0.00912501197308302, + 0.0038301139138638973, + 0.014170956797897816, + 0.0036860608961433172, + 0.003063264535740018, + 0.004084237385541201, + 0.01902635022997856, + 0.02393939532339573, + 0.009829962626099586, + 0.007810467854142189, + 0.03004699945449829, + 0.011477029882371426, + 0.006557755637913942, + 0.008812475949525833, + 0.003395763458684087, + 0.0030243468936532736, + 0.0029901701491326094, + 0.002108953194692731, + 0.004916978534311056, + 0.004579443950206041, + 0.0017428422579541802, + 0.0021967508364468813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.12714149057865143 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761949419975281 + ], + "opposite_logprob_tokens": [ + -0.8261949419975281 + ], + "logprob_reward": -0.5761949419975281, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761949419975281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31845200061798096, + 0.14603452384471893, + 0.03557860478758812, + 0.03327371925115585, + 0.08097280561923981, + 0.036519378423690796, + 0.07723948359489441, + 0.024894503876566887, + 0.18434257805347443, + 0.06269244104623795 + ], + "credit_map": [ + 0.04938937723636627, + 0.03369072079658508, + 0.04310803487896919, + 0.03510500118136406, + 0.054372694343328476, + 0.03987543657422066, + 0.026712127029895782, + 0.03619858995079994, + 0.029973220080137253, + 0.019327865913510323, + 0.01493509579449892, + 0.014270639978349209, + 0.02052338980138302, + 0.015387310646474361, + 0.017024412751197815, + 0.014592583291232586, + 0.0038898431230336428, + 0.00435452489182353, + 0.003164713503792882, + 0.005109788849949837, + 0.006403380539268255, + 0.003908089827746153, + 0.004678776487708092, + 0.004069488029927015, + 0.00994587130844593, + 0.004334554076194763, + 0.003776305355131626, + 0.002904506167396903, + 0.004199964925646782, + 0.002583895344287157, + 0.0028715950902551413, + 0.00265702698379755, + 0.018147747963666916, + 0.011667003855109215, + 0.007882139645516872, + 0.009173830971121788, + 0.009526167064905167, + 0.011972521431744099, + 0.0065379273146390915, + 0.006065470166504383, + 0.005207836627960205, + 0.0061571053229272366, + 0.003589895088225603, + 0.0042250496335327625, + 0.004854192491620779, + 0.006351980846375227, + 0.002864199923351407, + 0.0032691187225282192, + 0.011079232208430767, + 0.01191471703350544, + 0.007896407507359982, + 0.009387167170643806, + 0.010177452117204666, + 0.00917806662619114, + 0.010440720245242119, + 0.007165718823671341, + 0.0019353550160303712, + 0.004697533790022135, + 0.005374681670218706, + 0.0018542407779023051, + 0.0026375276502221823, + 0.003578243777155876, + 0.0022551771253347397, + 0.002561745699495077, + 0.024445604532957077, + 0.037961963564157486, + 0.015195967629551888, + 0.012957828119397163, + 0.04869262874126434, + 0.018779566511511803, + 0.010357808321714401, + 0.015951208770275116, + 0.013152310624718666, + 0.008925422094762325, + 0.005921982694417238, + 0.0036848874296993017, + 0.011837284080684185, + 0.009990233927965164, + 0.0037374980747699738, + 0.005442820489406586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.5761949419975281 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371884673833847 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371884673833847, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371884673833847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21764031052589417, + 0.17817004024982452, + 0.1455213874578476, + 0.05367221683263779, + 0.08082554489374161, + 0.0704205259680748, + 0.09845222532749176, + 0.043739203363657, + 0.0899629220366478, + 0.021595576778054237 + ], + "credit_map": [ + 0.0349222794175148, + 0.03419477492570877, + 0.018113544210791588, + 0.02053152024745941, + 0.037775006145238876, + 0.0330035462975502, + 0.018315354362130165, + 0.020784294232726097, + 0.018575454130768776, + 0.019963404163718224, + 0.014955426566302776, + 0.015131719410419464, + 0.04904117435216904, + 0.029371382668614388, + 0.013289612717926502, + 0.017841896042227745, + 0.014807069674134254, + 0.03457983210682869, + 0.009400896728038788, + 0.007754912134259939, + 0.04265367612242699, + 0.01595417968928814, + 0.008923054672777653, + 0.011447788216173649, + 0.010237306356430054, + 0.00733064953237772, + 0.005848649423569441, + 0.004424893297255039, + 0.007774997502565384, + 0.007343841250985861, + 0.005512588191777468, + 0.00519930012524128, + 0.01882263273000717, + 0.012747975997626781, + 0.0076927971094846725, + 0.007335218135267496, + 0.01058973465114832, + 0.011496033519506454, + 0.006130495574325323, + 0.006010666023939848, + 0.01760871335864067, + 0.009991651400923729, + 0.007245468907058239, + 0.008124706335365772, + 0.008644754998385906, + 0.007945031858980656, + 0.005674369633197784, + 0.005185834597796202, + 0.01419022586196661, + 0.014768218621611595, + 0.009525459259748459, + 0.011706795543432236, + 0.015245863236486912, + 0.009504871442914009, + 0.015312890522181988, + 0.008197905495762825, + 0.00465718237683177, + 0.005327396560460329, + 0.005812319461256266, + 0.004592567216604948, + 0.0067319441586732864, + 0.0058895377442240715, + 0.005089856684207916, + 0.005638402421027422, + 0.014573801308870316, + 0.014995202422142029, + 0.020719235762953758, + 0.00593387708067894, + 0.009370322339236736, + 0.00947808101773262, + 0.006313059478998184, + 0.008579343557357788, + 0.0028307668399065733, + 0.003113591345027089, + 0.0019988452550023794, + 0.0026107991579920053, + 0.00515822833403945, + 0.0030299504287540913, + 0.0014978188555687666, + 0.001355579006485641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03371884673833847 + }, + { + "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 high-angle drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldiers wearing uniform iron armor with a cold metallic sheen and peaked helmets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the soldiers' heavy boots tread in unison on dry, cracked earth, stirring up yellow-brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a glaring midday sun creating a grainy, highly realistic quality?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707560300827026 + ], + "opposite_logprob_tokens": [ + -1.137075662612915 + ], + "logprob_reward": -0.38707560300827026, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38707560300827026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11807519197463989, + 0.1414433717727661, + 0.12868517637252808, + 0.020295854657888412, + 0.047142598778009415, + 0.11194678395986557, + 0.09319742023944855, + 0.0413370318710804, + 0.142078697681427, + 0.15579786896705627 + ], + "credit_map": [ + 0.019692542031407356, + 0.012377356179058552, + 0.015586212277412415, + 0.01162164006382227, + 0.018117453902959824, + 0.013843236491084099, + 0.012240150943398476, + 0.0145966075360775, + 0.03380093723535538, + 0.020041650161147118, + 0.014715934172272682, + 0.012467220425605774, + 0.013418896123766899, + 0.01358197070658207, + 0.017204219475388527, + 0.016212541610002518, + 0.009278184734284878, + 0.03683985769748688, + 0.005609643645584583, + 0.005680733360350132, + 0.047316912561655045, + 0.01085072010755539, + 0.006215010304003954, + 0.006894116289913654, + 0.004713509697467089, + 0.0028511963319033384, + 0.002045469358563423, + 0.0016803297912701964, + 0.0030151589307934046, + 0.002567337127402425, + 0.001530266716144979, + 0.0018925870535895228, + 0.010137500241398811, + 0.007054501678794622, + 0.006473253946751356, + 0.0035084288101643324, + 0.005644611548632383, + 0.006050153635442257, + 0.004287589341402054, + 0.003986557479947805, + 0.029090167954564095, + 0.016372740268707275, + 0.011772195808589458, + 0.013318019919097424, + 0.012744411826133728, + 0.01255075354129076, + 0.008731456473469734, + 0.007367034908384085, + 0.014276392757892609, + 0.014772123657166958, + 0.010219743475317955, + 0.010073022916913033, + 0.0151752270758152, + 0.010144592262804508, + 0.01094296295195818, + 0.007593358401209116, + 0.005831858143210411, + 0.004133536946028471, + 0.0048303017392754555, + 0.0046242098324000835, + 0.005542630795389414, + 0.007274947129189968, + 0.0044858576729893684, + 0.004613686818629503, + 0.02098996564745903, + 0.028775718063116074, + 0.01226065680384636, + 0.009852861985564232, + 0.03454205393791199, + 0.015484350733458996, + 0.008764266967773438, + 0.011408821679651737, + 0.03481568023562431, + 0.02435627207159996, + 0.011905806139111519, + 0.00902594905346632, + 0.02906321920454502, + 0.02200653962790966, + 0.00959657970815897, + 0.015027835965156555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.38707560300827026 + } + ], + "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/0000155/group_0/index_tensor([8512, 8512])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_tensor([8512, 8512])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..acdb0b53ef9f31eb7efc8139db9de1975fd6f24b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_tensor([8512, 8512])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up, shot from eye level, focuses on an ear covered in dense, soft, light-brown hair. The outline of the auricle is clearly visible against a soft halo of light, a deep, blurred background of space. A thin, golden needle, gleaming with a brilliant metallic sheen, is carefully and steadily pulled outward by an invisible force. Its mirror-smooth shaft slowly slides out of the ear canal, its warm golden light illuminating the tips of every hair around it. As the needle completely exits the ear canal, the camera smoothly pulls back to reveal a wider scene. Simultaneously, a dazzling golden light erupts from the needle. Against the dark background, the needle expands and thickens at a visible speed, quickly transforming into a majestic staff with heavy golden hoops at each end and a deep ebony hue. A restrained yet powerful energy halo flows on the surface of the stick, casting dynamic light and shadow in the darkness. The entire picture transforms from the initial meticulousness to a grand narrative, full of mystery, power and a sense of epic that is about to unfold.", + "seed": 14125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_8512_seed_14125_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": -0.017567677795886992, + "vlm_logprob": -0.017567677795886992, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.017567677795886992, + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.012708628550171852 + ], + "opposite_logprob_tokens": [ + -5.01270866394043 + ], + "logprob_reward": -0.012708628550171852, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012708628550171852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7973039746284485, + 0.02199731022119522, + 0.029018929228186607, + 0.03140402212738991, + 0.01347985677421093, + 0.019642924889922142, + 0.015537751838564873, + 0.027218014001846313, + 0.027238601818680763, + 0.017158566042780876 + ], + "credit_map": [ + 0.05552782490849495, + 0.10104680806398392, + 0.2793921232223511, + 0.02574571967124939, + 0.048482634127140045, + 0.22193226218223572, + 0.020672034472227097, + 0.044504549354314804, + 0.002912837080657482, + 0.0029748729430139065, + 0.006071703042834997, + 0.00159456639084965, + 0.0017856124322861433, + 0.0016131229931488633, + 0.00170615641400218, + 0.0033384363632649183, + 0.003615122754126787, + 0.00513256061822176, + 0.007054531946778297, + 0.0022043867502361536, + 0.0028138950001448393, + 0.002280917251482606, + 0.00253146979957819, + 0.0033860416151583195, + 0.0029537088703364134, + 0.003002625424414873, + 0.010900505818426609, + 0.001913572079502046, + 0.002253751503303647, + 0.003232294926419854, + 0.0015443963930010796, + 0.005603164434432983, + 0.001773848314769566, + 0.0013792538084089756, + 0.0023340745829045773, + 0.0015110857784748077, + 0.0020617402624338865, + 0.0016104584792628884, + 0.0013164323754608631, + 0.0014929615426808596, + 0.003536636009812355, + 0.0018238984048366547, + 0.0024185627698898315, + 0.0023343798238784075, + 0.0024640453048050404, + 0.0020636445842683315, + 0.0023241559974849224, + 0.0026776003651320934, + 0.0021588322706520557, + 0.0016380955930799246, + 0.001867426559329033, + 0.0015533262630924582, + 0.001658548368141055, + 0.0013189426390454173, + 0.0028306690510362387, + 0.002511909930035472, + 0.0037572395522147417, + 0.003565155202522874, + 0.0022889284882694483, + 0.004546420648694038, + 0.002837487030774355, + 0.0020328746177256107, + 0.004382777493447065, + 0.0038071302697062492, + 0.005750979296863079, + 0.0016074858140200377, + 0.0019685456063598394, + 0.005496961995959282, + 0.002383531304076314, + 0.0027715363539755344, + 0.003960564266890287, + 0.0032989936880767345, + 0.002477051457390189, + 0.0022218639496713877, + 0.002478951821103692, + 0.002144162543118, + 0.0012713056057691574, + 0.0022018454037606716, + 0.0022431332617998123, + 0.0021202489733695984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012708628550171852 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.01608971692621708 + ], + "opposite_logprob_tokens": [ + -4.141089916229248 + ], + "logprob_reward": -0.01608971692621708, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608971692621708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34723979234695435, + 0.055713728070259094, + 0.026649950072169304, + 0.14531855285167694, + 0.07383910566568375, + 0.05764176696538925, + 0.07360377907752991, + 0.07997621595859528, + 0.09683767706155777, + 0.043179456144571304 + ], + "credit_map": [ + 0.02690873295068741, + 0.04843907803297043, + 0.12254281342029572, + 0.012425156310200691, + 0.0211348757147789, + 0.08749207854270935, + 0.010094703175127506, + 0.01820235513150692, + 0.01060235220938921, + 0.008485131897032261, + 0.011589440517127514, + 0.004872381221503019, + 0.003748165210708976, + 0.0043598320335149765, + 0.004343059379607439, + 0.007713362108916044, + 0.0029660617001354694, + 0.002802756614983082, + 0.004968229215592146, + 0.0019425612408667803, + 0.002558790845796466, + 0.004434431437402964, + 0.0034037854056805372, + 0.003573333378881216, + 0.012285304255783558, + 0.014895319007337093, + 0.05847112834453583, + 0.005745029076933861, + 0.00491815572604537, + 0.026089169085025787, + 0.0037031960673630238, + 0.019211232662200928, + 0.007938467897474766, + 0.007435645442456007, + 0.009735779836773872, + 0.01045033149421215, + 0.011854678392410278, + 0.008898599073290825, + 0.008345106616616249, + 0.009180494584143162, + 0.00757746072486043, + 0.0058720968663692474, + 0.005929385311901569, + 0.00725206034258008, + 0.005820737220346928, + 0.007196732796728611, + 0.011123408563435078, + 0.00686988141387701, + 0.012809782288968563, + 0.009640585631132126, + 0.00936343427747488, + 0.009420780465006828, + 0.007536093704402447, + 0.006080609746277332, + 0.008748254738748074, + 0.010004231706261635, + 0.007884763181209564, + 0.006231926381587982, + 0.005702013149857521, + 0.009471440687775612, + 0.007685021962970495, + 0.016214950010180473, + 0.014325636439025402, + 0.012460459023714066, + 0.011930643580853939, + 0.004829538986086845, + 0.007408569101244211, + 0.012825083918869495, + 0.011433803476393223, + 0.016792498528957367, + 0.018236946314573288, + 0.013380584307014942, + 0.0046884967014193535, + 0.0046220822259783745, + 0.004140207078307867, + 0.007072055246680975, + 0.003350032027810812, + 0.005152342841029167, + 0.006759787909686565, + 0.0073944516479969025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.01608971692621708 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.008646786212921143 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646786212921143, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646786212921143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5045794248580933, + 0.051477495580911636, + 0.020461535081267357, + 0.060226354748010635, + 0.061361923813819885, + 0.06824151426553726, + 0.04660589620471001, + 0.0472385510802269, + 0.09262976795434952, + 0.0471775084733963 + ], + "credit_map": [ + 0.039104070514440536, + 0.06103052198886871, + 0.17169514298439026, + 0.018242787569761276, + 0.031849827617406845, + 0.14088059961795807, + 0.014386685565114021, + 0.027389807626605034, + 0.008470936678349972, + 0.0074692838825285435, + 0.011941339820623398, + 0.003517782548442483, + 0.0023454031907022, + 0.0037150480784475803, + 0.0032697003334760666, + 0.010748004540801048, + 0.002597818849608302, + 0.00274648261256516, + 0.003885689890012145, + 0.0018141350010409951, + 0.0015467192279174924, + 0.0024995650164783, + 0.0016438426682725549, + 0.0037272830959409475, + 0.006498660426586866, + 0.007005638908594847, + 0.019982924684882164, + 0.003545548301190138, + 0.004474814981222153, + 0.004182768054306507, + 0.003292960114777088, + 0.011243042536079884, + 0.007418154738843441, + 0.005797042045742273, + 0.008087198249995708, + 0.009349219501018524, + 0.009617947041988373, + 0.006235136650502682, + 0.007427487522363663, + 0.0074297343380749226, + 0.006984023842960596, + 0.011548945680260658, + 0.009509003721177578, + 0.011336075142025948, + 0.008367800153791904, + 0.005956230219453573, + 0.006679947022348642, + 0.007859494537115097, + 0.009193039499223232, + 0.004737080540508032, + 0.004058795515447855, + 0.0053196679800748825, + 0.0052537438459694386, + 0.004322950262576342, + 0.006860693451017141, + 0.006859928835183382, + 0.005600298754870892, + 0.003798233810812235, + 0.0035060876980423927, + 0.004593576770275831, + 0.006534889806061983, + 0.005768380127847195, + 0.009247591719031334, + 0.0081894900649786, + 0.014759150333702564, + 0.007852070033550262, + 0.008225694298744202, + 0.011649860069155693, + 0.009475107304751873, + 0.00967917125672102, + 0.015656672418117523, + 0.015332038514316082, + 0.0066342600621283054, + 0.004092566203325987, + 0.004559015855193138, + 0.006510759238153696, + 0.004343231674283743, + 0.006435375194996595, + 0.008597747422754765, + 0.006004551891237497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.008646786212921143 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.029809197410941124 + ], + "opposite_logprob_tokens": [ + -3.5298092365264893 + ], + "logprob_reward": -0.029809197410941124, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809197410941124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5062556266784668, + 0.05305032059550285, + 0.03462526574730873, + 0.04162516072392464, + 0.040791116654872894, + 0.054971568286418915, + 0.024389585480093956, + 0.07698047906160355, + 0.09850986301898956, + 0.06880103796720505 + ], + "credit_map": [ + 0.0431935116648674, + 0.06339643895626068, + 0.18699999153614044, + 0.01831168867647648, + 0.029420185834169388, + 0.12407835572957993, + 0.015234611928462982, + 0.025620801374316216, + 0.009748567827045918, + 0.0072123645804822445, + 0.010564579628407955, + 0.0037083690986037254, + 0.004542567301541567, + 0.0038860307540744543, + 0.0038896265905350447, + 0.009498214349150658, + 0.004402637481689453, + 0.0057191201485693455, + 0.00837674830108881, + 0.002745718462392688, + 0.0027065768372267485, + 0.0029380284249782562, + 0.0029212050139904022, + 0.004815230146050453, + 0.006181852892041206, + 0.0056958189234137535, + 0.010384632274508476, + 0.003662422765046358, + 0.00373569643124938, + 0.004352020099759102, + 0.0026434333994984627, + 0.004969284404069185, + 0.005946798715740442, + 0.003190165152773261, + 0.007900157943367958, + 0.005639954004436731, + 0.003951160702854395, + 0.003996134735643864, + 0.004764362704008818, + 0.005402382463216782, + 0.00578164542093873, + 0.004969228059053421, + 0.006436578463762999, + 0.007741122040897608, + 0.008830845355987549, + 0.005918148439377546, + 0.008267197757959366, + 0.007026804145425558, + 0.0038634033408015966, + 0.002265059854835272, + 0.0022724478039890528, + 0.00275783846154809, + 0.002239648485556245, + 0.0026977478992193937, + 0.004876232706010342, + 0.0034172069281339645, + 0.009024174883961678, + 0.005680151749402285, + 0.005367664620280266, + 0.010339316911995411, + 0.006750475149601698, + 0.012645473703742027, + 0.0148472860455513, + 0.01232594158500433, + 0.020362786948680878, + 0.00640106899663806, + 0.007413818966597319, + 0.01680813916027546, + 0.008721437305212021, + 0.010144907981157303, + 0.015467891469597816, + 0.013189811259508133, + 0.007746290415525436, + 0.006403601728379726, + 0.00718979025259614, + 0.0079411705955863, + 0.006365915294736624, + 0.009852474555373192, + 0.011018713004887104, + 0.012283078394830227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.029809197410941124 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.02058405987918377 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058405987918377, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058405987918377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14542728662490845, + 0.03591011092066765, + 0.06664836406707764, + 0.07133901864290237, + 0.0951392725110054, + 0.1546175330877304, + 0.10106805711984634, + 0.04585017263889313, + 0.14868025481700897, + 0.13531994819641113 + ], + "credit_map": [ + 0.019661054015159607, + 0.016079766675829887, + 0.04519663378596306, + 0.00842269416898489, + 0.011397236958146095, + 0.019231289625167847, + 0.010955089703202248, + 0.014483524486422539, + 0.0039022155106067657, + 0.004664559382945299, + 0.008113153278827667, + 0.002465021563693881, + 0.003527082735672593, + 0.003462280612438917, + 0.003267101012170315, + 0.0065086958929896355, + 0.006327912677079439, + 0.008196580223739147, + 0.012872411869466305, + 0.004048698116093874, + 0.005092032719403505, + 0.014397739432752132, + 0.00542647298425436, + 0.010286517441272736, + 0.007953786291182041, + 0.007463508751243353, + 0.009482610039412975, + 0.004211672116070986, + 0.0036978572607040405, + 0.01565033756196499, + 0.0029948456212878227, + 0.019884400069713593, + 0.011464567855000496, + 0.011874875985085964, + 0.016839727759361267, + 0.013681999407708645, + 0.013202715665102005, + 0.007516241632401943, + 0.00856532622128725, + 0.011993815191090107, + 0.01737220399081707, + 0.015515396371483803, + 0.016893427819013596, + 0.0223991796374321, + 0.021637221798300743, + 0.016225019469857216, + 0.02443954162299633, + 0.020135536789894104, + 0.016705097630620003, + 0.006624302826821804, + 0.007426762953400612, + 0.01458356436342001, + 0.01135371532291174, + 0.008622877299785614, + 0.020891526713967323, + 0.014860200695693493, + 0.0057183499448001385, + 0.004560266621410847, + 0.003550820518285036, + 0.006352717988193035, + 0.003985876217484474, + 0.00583343580365181, + 0.008634722791612148, + 0.007213985081762075, + 0.015141254290938377, + 0.0075220526196062565, + 0.009673149324953556, + 0.018011189997196198, + 0.012093071825802326, + 0.030889656394720078, + 0.0268124807626009, + 0.028537394478917122, + 0.016882600262761116, + 0.007798709440976381, + 0.008693336509168148, + 0.01979134976863861, + 0.009215475991368294, + 0.022830838337540627, + 0.025498345494270325, + 0.024609295651316643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.02058405987918377 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, shot from eye level, focuses on an ear covered in dense, soft, light-brown hair. The outline of the auricle is clearly visible against a soft halo of light, a deep, blurred background of space. A thin, golden needle, gleaming with a brilliant metallic sheen, is carefully and steadily pulled outward by an invisible force. Its mirror-smooth shaft slowly slides out of the ear canal, its warm golden light illuminating the tips of every hair around it. As the needle completely exits the ear canal, the camera smoothly pulls back to reveal a wider scene. Simultaneously, a dazzling golden light erupts from the needle. Against the dark background, the needle expands and thickens at a visible speed, quickly transforming into a majestic staff with heavy golden hoops at each end and a deep ebony hue. A restrained yet powerful energy halo flows on the surface of the stick, casting dynamic light and shadow in the darkness. The entire picture transforms from the initial meticulousness to a grand narrative, full of mystery, power and a sense of epic that is about to unfold.", + "seed": 14126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_8512_seed_14126_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": -0.02117433687672019, + "vlm_logprob": -0.02117433687672019, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02117433687672019, + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.01575099676847458 + ], + "opposite_logprob_tokens": [ + -4.640750885009766 + ], + "logprob_reward": -0.01575099676847458, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01575099676847458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17518846690654755, + 0.041685156524181366, + 0.10188072919845581, + 0.14237406849861145, + 0.05548219010233879, + 0.0709194540977478, + 0.0937788262963295, + 0.14821180701255798, + 0.11645606905221939, + 0.054023224860429764 + ], + "credit_map": [ + 0.01781538687646389, + 0.021368473768234253, + 0.023758305236697197, + 0.029977304860949516, + 0.023662567138671875, + 0.026294268667697906, + 0.013220234774053097, + 0.019091930240392685, + 0.0052777547389268875, + 0.005369739141315222, + 0.005197112914174795, + 0.005762866698205471, + 0.005011934321373701, + 0.005433143116533756, + 0.004058042075484991, + 0.00557456212118268, + 0.01064968854188919, + 0.016484173014760017, + 0.01415154431015253, + 0.013898195698857307, + 0.01491217128932476, + 0.009093103930354118, + 0.013840761967003345, + 0.008851082064211369, + 0.01093708910048008, + 0.008601272478699684, + 0.04898000881075859, + 0.049244582653045654, + 0.005508460570126772, + 0.004474211949855089, + 0.007529199123382568, + 0.007099235896021128, + 0.00428011454641819, + 0.006295592524111271, + 0.013299793004989624, + 0.016159188002347946, + 0.0038939642254263163, + 0.0032157094683498144, + 0.003992966841906309, + 0.004344858694821596, + 0.011009673587977886, + 0.008446307852864265, + 0.015525100752711296, + 0.013137271627783775, + 0.00610358826816082, + 0.005143526941537857, + 0.006729003041982651, + 0.004824984818696976, + 0.007289784960448742, + 0.010643327608704567, + 0.029458366334438324, + 0.030166447162628174, + 0.004882441833615303, + 0.0030777226202189922, + 0.004912943579256535, + 0.0033477842807769775, + 0.030270732939243317, + 0.021330147981643677, + 0.019629890099167824, + 0.02284074015915394, + 0.013550996780395508, + 0.012580725364387035, + 0.01823270507156849, + 0.009775867685675621, + 0.008797841146588326, + 0.0292294230312109, + 0.014401838183403015, + 0.025534523651003838, + 0.018251841887831688, + 0.006279196124523878, + 0.00816968735307455, + 0.005791721865534782, + 0.007014359813183546, + 0.005534098483622074, + 0.007261960301548243, + 0.008744105696678162, + 0.0059618111699819565, + 0.004485557321459055, + 0.010581678710877895, + 0.004439658019691706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.01575099676847458 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.026371391490101814 + ], + "opposite_logprob_tokens": [ + -3.651371479034424 + ], + "logprob_reward": -0.026371391490101814, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371391490101814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07490904629230499, + 0.09971964359283447, + 0.0598558709025383, + 0.1683899462223053, + 0.11579787731170654, + 0.1345374435186386, + 0.11914432793855667, + 0.08798781037330627, + 0.06451533734798431, + 0.07514265179634094 + ], + "credit_map": [ + 0.010618390515446663, + 0.00792586337774992, + 0.009711683727800846, + 0.009057675488293171, + 0.009472365491092205, + 0.007025007624179125, + 0.01003935094922781, + 0.01105869933962822, + 0.008718344382941723, + 0.01005459576845169, + 0.011023112572729588, + 0.009433621540665627, + 0.019688362255692482, + 0.015253771096467972, + 0.010579428635537624, + 0.014968409202992916, + 0.0050851269625127316, + 0.006952223833650351, + 0.008267045952379704, + 0.008329602889716625, + 0.010568428784608841, + 0.008889582008123398, + 0.005738255102187395, + 0.006025606766343117, + 0.009299902245402336, + 0.007713595870882273, + 0.06352631002664566, + 0.06682053208351135, + 0.0037932484410703182, + 0.0045049721375107765, + 0.005447321105748415, + 0.0072840740904212, + 0.0074783493764698505, + 0.013221761211752892, + 0.021155240014195442, + 0.023182008415460587, + 0.01239862758666277, + 0.013273406773805618, + 0.010950727388262749, + 0.0141377542167902, + 0.018811311572790146, + 0.01676945388317108, + 0.02397705800831318, + 0.027074597775936127, + 0.012810343876481056, + 0.009191036224365234, + 0.014732996933162212, + 0.011170644313097, + 0.01200633030384779, + 0.007581463549286127, + 0.03193886950612068, + 0.03583644703030586, + 0.005548735614866018, + 0.007674309890717268, + 0.01094333827495575, + 0.007614833302795887, + 0.016292409971356392, + 0.010023812763392925, + 0.012470259331166744, + 0.01330280490219593, + 0.011138515546917915, + 0.007989087142050266, + 0.011354025453329086, + 0.005416900385171175, + 0.005314957816153765, + 0.016599280759692192, + 0.0071182833053171635, + 0.016908839344978333, + 0.005976288579404354, + 0.0032395964954048395, + 0.006699661258608103, + 0.0026584307197481394, + 0.006213122047483921, + 0.011734060011804104, + 0.01897333562374115, + 0.011374996975064278, + 0.009609404951334, + 0.004353753291070461, + 0.009624093770980835, + 0.0032598807010799646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.026371391490101814 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.00674506975337863 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.00674506975337863, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674506975337863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045672085136175156, + 0.04027000442147255, + 0.06842686235904694, + 0.08267296850681305, + 0.09705117344856262, + 0.11713607609272003, + 0.14870385825634003, + 0.1923597753047943, + 0.1380866914987564, + 0.06962043792009354 + ], + "credit_map": [ + 0.0064133452251553535, + 0.0047844573855400085, + 0.006271907594054937, + 0.005022687371820211, + 0.006507639307528734, + 0.005875342059880495, + 0.0054041496478021145, + 0.005392555613070726, + 0.0038732532411813736, + 0.004772312007844448, + 0.005725950468331575, + 0.007645241916179657, + 0.004213364329189062, + 0.005334631074219942, + 0.004261689726263285, + 0.004443560726940632, + 0.005386925768107176, + 0.011362064629793167, + 0.01000829879194498, + 0.009202130138874054, + 0.009112942963838577, + 0.006708316970616579, + 0.010604853741824627, + 0.006041334941983223, + 0.005376321263611317, + 0.005522394552826881, + 0.025408213958144188, + 0.02791234664618969, + 0.00380523968487978, + 0.00334391207434237, + 0.006566326133906841, + 0.004738218151032925, + 0.006596886552870274, + 0.008640454150736332, + 0.01709803380072117, + 0.018081676214933395, + 0.011493216268718243, + 0.010287811048328876, + 0.011402804404497147, + 0.01345028541982174, + 0.02170911617577076, + 0.013713506050407887, + 0.026377661153674126, + 0.025712477043271065, + 0.006352952681481838, + 0.00581824267283082, + 0.009745432995259762, + 0.007706694304943085, + 0.022253261879086494, + 0.017261207103729248, + 0.04182516783475876, + 0.039578866213560104, + 0.008753422647714615, + 0.005825067404657602, + 0.007980037480592728, + 0.005226833280175924, + 0.022374259307980537, + 0.0328335165977478, + 0.020864948630332947, + 0.030261700972914696, + 0.027433229610323906, + 0.016309797763824463, + 0.026295404881238937, + 0.01598692685365677, + 0.011447842232882977, + 0.03654422238469124, + 0.016169875860214233, + 0.03221987932920456, + 0.017608225345611572, + 0.0070687104016542435, + 0.00984505470842123, + 0.007182871922850609, + 0.007089171092957258, + 0.012094699777662754, + 0.010313586331903934, + 0.013073699548840523, + 0.008305700495839119, + 0.007380684372037649, + 0.007910653948783875, + 0.0034522442147135735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.00674506975337863 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.03370916470885277 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.03370916470885277, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370916470885277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028883911669254303, + 0.023783963173627853, + 0.038923971354961395, + 0.3206620216369629, + 0.059235602617263794, + 0.26922303438186646, + 0.07386334985494614, + 0.06258224695920944, + 0.07658874243497849, + 0.04625320807099342 + ], + "credit_map": [ + 0.003321168478578329, + 0.004046532791107893, + 0.003879118012264371, + 0.0034207634162157774, + 0.0038484518881887197, + 0.0042677330784499645, + 0.0031934217549860477, + 0.0029067236464470625, + 0.0024077214766293764, + 0.00291338749229908, + 0.0037202679086476564, + 0.005070074461400509, + 0.002549757482483983, + 0.0019568903371691704, + 0.0022672261111438274, + 0.0028986369725316763, + 0.003406570525839925, + 0.004849676508456469, + 0.005419968161731958, + 0.006222221069037914, + 0.005767103750258684, + 0.0036675448063760996, + 0.005602429620921612, + 0.003988457378000021, + 0.015701429918408394, + 0.02654896304011345, + 0.12260609120130539, + 0.12916357815265656, + 0.005873208865523338, + 0.003537832759320736, + 0.010295453481376171, + 0.00693544652312994, + 0.006352444179356098, + 0.0049473256804049015, + 0.012403193861246109, + 0.013196783140301704, + 0.005450135096907616, + 0.004376339726150036, + 0.005836808122694492, + 0.006672571878880262, + 0.05915522202849388, + 0.043140437453985214, + 0.054278548806905746, + 0.047758616507053375, + 0.01530357077717781, + 0.011393495835363865, + 0.024915875867009163, + 0.013277268037199974, + 0.008333941921591759, + 0.009955188259482384, + 0.01881301775574684, + 0.02047078125178814, + 0.005156587343662977, + 0.00328832701779902, + 0.004272960126399994, + 0.003572544315829873, + 0.011008565314114094, + 0.009332440793514252, + 0.012638440355658531, + 0.011422795243561268, + 0.007030852138996124, + 0.0038256109692156315, + 0.004269039258360863, + 0.003054505679756403, + 0.008150169625878334, + 0.015023420564830303, + 0.0153388362377882, + 0.01316907163709402, + 0.012198922224342823, + 0.0037357325199991465, + 0.005441464949399233, + 0.003531123511493206, + 0.0036547514609992504, + 0.004809074103832245, + 0.007702372502535582, + 0.008711041882634163, + 0.0044368174858391285, + 0.00785648450255394, + 0.004997787531465292, + 0.004084879532456398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.03370916470885277 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.02329506166279316 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329506166279316, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329506166279316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031661100685596466, + 0.04173280671238899, + 0.017815450206398964, + 0.10604295134544373, + 0.0510815754532814, + 0.03443609178066254, + 0.04669864475727081, + 0.09126953780651093, + 0.43622973561286926, + 0.14303210377693176 + ], + "credit_map": [ + 0.0027601104229688644, + 0.003915772307664156, + 0.0039975144900381565, + 0.004434862174093723, + 0.004351489245891571, + 0.005187567323446274, + 0.0034131016582250595, + 0.0036006851587444544, + 0.004603947978466749, + 0.0064936913549900055, + 0.0058968933299183846, + 0.008208146318793297, + 0.003604958765208721, + 0.003981945104897022, + 0.003932139370590448, + 0.00501108868047595, + 0.0015159330796450377, + 0.0023697030264884233, + 0.002489023841917515, + 0.002615712583065033, + 0.0023016564082354307, + 0.0019137527560815215, + 0.0027990289963781834, + 0.0018106400966644287, + 0.003986232448369265, + 0.008236702531576157, + 0.040112487971782684, + 0.04355498030781746, + 0.0022091364953666925, + 0.0015061706071719527, + 0.0037023501936346292, + 0.0027348962612450123, + 0.0029078074730932713, + 0.00411771098151803, + 0.008431965485215187, + 0.005770595744252205, + 0.006793753243982792, + 0.0063233934342861176, + 0.0076814135536551476, + 0.009054942056536674, + 0.00716890674084425, + 0.004738833289593458, + 0.006187319755554199, + 0.006182962097227573, + 0.0027907048352062702, + 0.002007547300308943, + 0.003071722574532032, + 0.0022880982141941786, + 0.00604519946500659, + 0.004970923997461796, + 0.009764040820300579, + 0.011213816702365875, + 0.0033111339434981346, + 0.003035794710740447, + 0.004736158065497875, + 0.0036215789150446653, + 0.027406495064496994, + 0.011609832756221294, + 0.014352391473948956, + 0.01550835557281971, + 0.009395463392138481, + 0.004353211726993322, + 0.004819937516003847, + 0.0038238551933318377, + 0.04693221300840378, + 0.08165768533945084, + 0.08781696110963821, + 0.06737664341926575, + 0.07263229787349701, + 0.021342583000659943, + 0.03559669852256775, + 0.02287464775145054, + 0.01563161611557007, + 0.028716644272208214, + 0.03264012560248375, + 0.01708752103149891, + 0.025795655325055122, + 0.006095213815569878, + 0.013116398826241493, + 0.003948943689465523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.02329506166279316 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, shot from eye level, focuses on an ear covered in dense, soft, light-brown hair. The outline of the auricle is clearly visible against a soft halo of light, a deep, blurred background of space. A thin, golden needle, gleaming with a brilliant metallic sheen, is carefully and steadily pulled outward by an invisible force. Its mirror-smooth shaft slowly slides out of the ear canal, its warm golden light illuminating the tips of every hair around it. As the needle completely exits the ear canal, the camera smoothly pulls back to reveal a wider scene. Simultaneously, a dazzling golden light erupts from the needle. Against the dark background, the needle expands and thickens at a visible speed, quickly transforming into a majestic staff with heavy golden hoops at each end and a deep ebony hue. A restrained yet powerful energy halo flows on the surface of the stick, casting dynamic light and shadow in the darkness. The entire picture transforms from the initial meticulousness to a grand narrative, full of mystery, power and a sense of epic that is about to unfold.", + "seed": 14127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_8512_seed_14127_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": -0.02818683385848999, + "vlm_logprob": -0.02818683385848999, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.02818683385848999, + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.018465986475348473 + ], + "opposite_logprob_tokens": [ + -4.643465995788574 + ], + "logprob_reward": -0.018465986475348473, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018465986475348473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18317227065563202, + 0.02731381170451641, + 0.028520137071609497, + 0.04857439920306206, + 0.19721169769763947, + 0.1919833868741989, + 0.1099112331867218, + 0.08866959065198898, + 0.044728197157382965, + 0.07991521060466766 + ], + "credit_map": [ + 0.027187950909137726, + 0.02560340240597725, + 0.026760002598166466, + 0.015555597841739655, + 0.014458823017776012, + 0.027875257655978203, + 0.021448835730552673, + 0.024282407015562057, + 0.004273313097655773, + 0.002111593494191766, + 0.004671395290642977, + 0.001660134643316269, + 0.002061333041638136, + 0.005515269935131073, + 0.002406924031674862, + 0.004613853059709072, + 0.003528321161866188, + 0.002680748701095581, + 0.004008791409432888, + 0.0022034484427422285, + 0.0024233211297541857, + 0.00516691617667675, + 0.0034041120670735836, + 0.00510448170825839, + 0.0058507188223302364, + 0.0034843480680137873, + 0.007346696685999632, + 0.003844636492431164, + 0.002910642186179757, + 0.010137286968529224, + 0.004452917259186506, + 0.0105471545830369, + 0.02879113145172596, + 0.004374665208160877, + 0.011707872152328491, + 0.004896615631878376, + 0.0061618490144610405, + 0.06872960180044174, + 0.007602188270539045, + 0.06494780629873276, + 0.060614198446273804, + 0.00963753554970026, + 0.03464122489094734, + 0.006536158267408609, + 0.007496463134884834, + 0.03203964605927467, + 0.00929830502718687, + 0.0317198783159256, + 0.013576701283454895, + 0.008438029326498508, + 0.009810879826545715, + 0.005740168504416943, + 0.007335431408137083, + 0.027626803144812584, + 0.010536528192460537, + 0.02684669941663742, + 0.00815805234014988, + 0.006628187373280525, + 0.007249080576002598, + 0.008333354257047176, + 0.00626784423366189, + 0.023792190477252007, + 0.0066321068443357944, + 0.021608782932162285, + 0.0048405881971120834, + 0.005550819914788008, + 0.007563687860965729, + 0.005567921791225672, + 0.005770126357674599, + 0.0056030722334980965, + 0.004355758428573608, + 0.00547622749581933, + 0.00868243258446455, + 0.008351458236575127, + 0.010574535466730595, + 0.0072995624504983425, + 0.01551466342061758, + 0.010571521706879139, + 0.009883643127977848, + 0.009037399664521217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.018465986475348473 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.048696789890527725 + ], + "opposite_logprob_tokens": [ + -3.048696756362915 + ], + "logprob_reward": -0.048696789890527725, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048696789890527725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06592392921447754, + 0.06544326990842819, + 0.020841924473643303, + 0.06546299904584885, + 0.08192360401153564, + 0.31888115406036377, + 0.08797916769981384, + 0.12600427865982056, + 0.10975953936576843, + 0.05778009444475174 + ], + "credit_map": [ + 0.0072799441404640675, + 0.007132616359740496, + 0.00887257419526577, + 0.004969129338860512, + 0.00723672891035676, + 0.010890658013522625, + 0.009563460014760494, + 0.009978813119232655, + 0.010860908776521683, + 0.007862035185098648, + 0.011578257195651531, + 0.005019635893404484, + 0.004806090146303177, + 0.009558893740177155, + 0.005835089832544327, + 0.009922357276082039, + 0.0035554394125938416, + 0.0020037954673171043, + 0.003066315781325102, + 0.0021305792033672333, + 0.0017260678578168154, + 0.0030145279597491026, + 0.001863190671429038, + 0.0034820083528757095, + 0.008948804810643196, + 0.003813050454482436, + 0.009848964400589466, + 0.0048577976413071156, + 0.00427434453740716, + 0.013727597892284393, + 0.007215949706733227, + 0.012776488438248634, + 0.01193985715508461, + 0.004808569792658091, + 0.006902736611664295, + 0.008982106111943722, + 0.0033624223433434963, + 0.02151617407798767, + 0.004409067332744598, + 0.020002668723464012, + 0.08744927495718002, + 0.019394980743527412, + 0.06247643753886223, + 0.013154993765056133, + 0.016223261132836342, + 0.05248796194791794, + 0.014517927542328835, + 0.05317629128694534, + 0.008212864398956299, + 0.004413422662764788, + 0.008602487854659557, + 0.0033861612901091576, + 0.00479422602802515, + 0.026625443249940872, + 0.007485059555619955, + 0.02445949986577034, + 0.008321316912770271, + 0.005201214924454689, + 0.008116393350064754, + 0.0059013064019382, + 0.003899004077538848, + 0.043110191822052, + 0.006390956696122885, + 0.04506390541791916, + 0.01860908418893814, + 0.011616387404501438, + 0.016145050525665283, + 0.011583278886973858, + 0.01323486678302288, + 0.01662089303135872, + 0.008865023031830788, + 0.01308495458215475, + 0.0076938895508646965, + 0.007348873186856508, + 0.0069593340158462524, + 0.00624253461137414, + 0.008448162116110325, + 0.007067433558404446, + 0.007529917638748884, + 0.0064899493008852005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.048696789890527725 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.012552540749311447 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552540749311447, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552540749311447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041951581835746765, + 0.030095752328634262, + 0.03876477852463722, + 0.08883919566869736, + 0.1505325436592102, + 0.10134795308113098, + 0.09849441796541214, + 0.14517056941986084, + 0.16443322598934174, + 0.1403699517250061 + ], + "credit_map": [ + 0.004308568313717842, + 0.0043699475936591625, + 0.006573623977601528, + 0.0034105039667338133, + 0.0032316932920366526, + 0.008230028674006462, + 0.004652564879506826, + 0.0071746548637747765, + 0.0047644563019275665, + 0.0025970954447984695, + 0.004817637149244547, + 0.0019479509210214019, + 0.0018663821974769235, + 0.00592360133305192, + 0.002575390739366412, + 0.005603244062513113, + 0.0055626980029046535, + 0.003714900463819504, + 0.004642360843718052, + 0.0032509153243154287, + 0.0030579050071537495, + 0.007846810854971409, + 0.0037058244924992323, + 0.006983367260545492, + 0.011944923549890518, + 0.004864667076617479, + 0.012579729780554771, + 0.006550133693963289, + 0.0056612323969602585, + 0.019063502550125122, + 0.008862735703587532, + 0.019312279298901558, + 0.019258039072155952, + 0.00575007451698184, + 0.010062948800623417, + 0.011907219886779785, + 0.006583457347005606, + 0.044875405728816986, + 0.0074222274124622345, + 0.04467318579554558, + 0.015605823136866093, + 0.007949599996209145, + 0.020163213834166527, + 0.006859538145363331, + 0.008111657574772835, + 0.018397657200694084, + 0.006935080513358116, + 0.017325391992926598, + 0.017258822917938232, + 0.00939891580492258, + 0.021493621170520782, + 0.008386828936636448, + 0.005082889925688505, + 0.016486963257193565, + 0.0061666835099458694, + 0.01421970222145319, + 0.01779956743121147, + 0.008828082121908665, + 0.01845628209412098, + 0.011149553582072258, + 0.008575434796512127, + 0.03480520099401474, + 0.009043197147548199, + 0.03651326522231102, + 0.022462302818894386, + 0.015597179532051086, + 0.019301068037748337, + 0.027568640187382698, + 0.02194177731871605, + 0.018981749191880226, + 0.01715410128235817, + 0.021426429972052574, + 0.01822969876229763, + 0.017120735719799995, + 0.01649102196097374, + 0.0140274278819561, + 0.02427210845053196, + 0.017989812418818474, + 0.01678367704153061, + 0.015455478802323341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.012552540749311447 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.04867090284824371 + ], + "opposite_logprob_tokens": [ + -3.048671007156372 + ], + "logprob_reward": -0.04867090284824371, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867090284824371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07834669947624207, + 0.03021995723247528, + 0.0302510354667902, + 0.053582027554512024, + 0.10525147616863251, + 0.22580765187740326, + 0.11138428747653961, + 0.17296923696994781, + 0.1035194844007492, + 0.08866813778877258 + ], + "credit_map": [ + 0.007433195598423481, + 0.008687133900821209, + 0.008800229988992214, + 0.006205401383340359, + 0.009178870357573032, + 0.014156144112348557, + 0.010874046944081783, + 0.013011683709919453, + 0.002831095829606056, + 0.0026447405107319355, + 0.0041788602247834206, + 0.002377380384132266, + 0.0028956138994544744, + 0.006270134821534157, + 0.003215628210455179, + 0.0058065070770680904, + 0.004392113070935011, + 0.0027883087750524282, + 0.0038910293951630592, + 0.00269132643006742, + 0.0023606782779097557, + 0.003960629925131798, + 0.004280198831111193, + 0.0058867535553872585, + 0.007774855941534042, + 0.00275941239669919, + 0.0085689602419734, + 0.004657046403735876, + 0.0033466212917119265, + 0.010499682277441025, + 0.0047836508601903915, + 0.01119180116802454, + 0.012627369724214077, + 0.00837024673819542, + 0.009473713114857674, + 0.019108960404992104, + 0.0076898797415196896, + 0.017763203009963036, + 0.01374499499797821, + 0.01647311821579933, + 0.07577023655176163, + 0.010099905543029308, + 0.039556749165058136, + 0.009920152835547924, + 0.009759278036653996, + 0.03617364168167114, + 0.012079150415956974, + 0.032448556274175644, + 0.016529658809304237, + 0.00802079401910305, + 0.020097265020012856, + 0.008154978044331074, + 0.004531501326709986, + 0.025413138791918755, + 0.00620258878916502, + 0.02243436500430107, + 0.010126819834113121, + 0.007654948625713587, + 0.01134723611176014, + 0.01397942565381527, + 0.008230402134358883, + 0.05647337809205055, + 0.008913910016417503, + 0.056243132799863815, + 0.01366337575018406, + 0.011003117077052593, + 0.014865659177303314, + 0.015079925768077374, + 0.011191318742930889, + 0.014912037178874016, + 0.010068973526358604, + 0.012735079042613506, + 0.010774198919534683, + 0.008986788801848888, + 0.013591863214969635, + 0.010319883935153484, + 0.0131752984598279, + 0.009923973120748997, + 0.01220802590250969, + 0.00968810822814703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.04867090284824371 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0713895857334137, + 0.026329584419727325, + 0.0321439690887928, + 0.05222191661596298, + 0.07887576520442963, + 0.15391914546489716, + 0.1335587352514267, + 0.06564104557037354, + 0.18914493918418884, + 0.19677527248859406 + ], + "credit_map": [ + 0.007890800014138222, + 0.00924740917980671, + 0.008861408568918705, + 0.005940884817391634, + 0.006496448069810867, + 0.014113495126366615, + 0.007986893877387047, + 0.01085224375128746, + 0.002683383645489812, + 0.001645475858822465, + 0.004627001937478781, + 0.001879658317193389, + 0.0021659533958882093, + 0.005329222418367863, + 0.0026536958757787943, + 0.0053451936691999435, + 0.004200432449579239, + 0.003035852685570717, + 0.005521279294043779, + 0.0026792511343955994, + 0.00244649569503963, + 0.005660722032189369, + 0.0035845215898007154, + 0.005015410948544741, + 0.009449910372495651, + 0.0026694906409829855, + 0.0073140873573720455, + 0.00468820845708251, + 0.003417834173887968, + 0.010997888632118702, + 0.004209662321954966, + 0.009474837221205235, + 0.011860274709761143, + 0.002101532882079482, + 0.004299826920032501, + 0.0039618718437850475, + 0.002284367335960269, + 0.025212856009602547, + 0.004878939129412174, + 0.024276096373796463, + 0.02339298464357853, + 0.013019315898418427, + 0.033760569989681244, + 0.011004836298525333, + 0.011959414929151535, + 0.02611931785941124, + 0.010348965413868427, + 0.024313755333423615, + 0.029869843274354935, + 0.009365318343043327, + 0.024664046242833138, + 0.01135818287730217, + 0.007780505809932947, + 0.022322140634059906, + 0.007899858988821507, + 0.020298844203352928, + 0.008026372641324997, + 0.005261773243546486, + 0.007979581132531166, + 0.005173258017748594, + 0.0038006643299013376, + 0.015487193129956722, + 0.00518413819372654, + 0.014728059992194176, + 0.030726423487067223, + 0.018114252015948296, + 0.023010840639472008, + 0.01884620264172554, + 0.02482474222779274, + 0.03279007971286774, + 0.016005635261535645, + 0.024826757609844208, + 0.017095360904932022, + 0.018179764971137047, + 0.0194062739610672, + 0.02071204222738743, + 0.04643355309963226, + 0.022490795701742172, + 0.02964841201901436, + 0.022809065878391266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.012547949329018593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up, shot from eye level, focuses on an ear covered in dense, soft, light-brown hair. The outline of the auricle is clearly visible against a soft halo of light, a deep, blurred background of space. A thin, golden needle, gleaming with a brilliant metallic sheen, is carefully and steadily pulled outward by an invisible force. Its mirror-smooth shaft slowly slides out of the ear canal, its warm golden light illuminating the tips of every hair around it. As the needle completely exits the ear canal, the camera smoothly pulls back to reveal a wider scene. Simultaneously, a dazzling golden light erupts from the needle. Against the dark background, the needle expands and thickens at a visible speed, quickly transforming into a majestic staff with heavy golden hoops at each end and a deep ebony hue. A restrained yet powerful energy halo flows on the surface of the stick, casting dynamic light and shadow in the darkness. The entire picture transforms from the initial meticulousness to a grand narrative, full of mystery, power and a sense of epic that is about to unfold.", + "seed": 14128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_0/index_8512_seed_14128_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": -0.028902792371809483, + "vlm_logprob": -0.028902792371809483, + "vlm_yes_no_margin": 3.974999952316284, + "vlm_token_credit": -0.028902792371809483, + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.016547108069062233 + ], + "opposite_logprob_tokens": [ + -4.641547203063965 + ], + "logprob_reward": -0.016547108069062233, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.016547108069062233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22334589064121246, + 0.0844365656375885, + 0.06323826313018799, + 0.10931168496608734, + 0.11028441041707993, + 0.09443206340074539, + 0.08607891201972961, + 0.07730384916067123, + 0.061753686517477036, + 0.0898146778345108 + ], + "credit_map": [ + 0.031497567892074585, + 0.023756088688969612, + 0.03483187407255173, + 0.02018764428794384, + 0.020389677956700325, + 0.026615925133228302, + 0.02279055491089821, + 0.04327655956149101, + 0.011605759151279926, + 0.00762623967602849, + 0.01158118899911642, + 0.008123781532049179, + 0.009324589744210243, + 0.009778370149433613, + 0.013739846646785736, + 0.012656783685088158, + 0.0076178922317922115, + 0.003811068832874298, + 0.008431889116764069, + 0.006586665287613869, + 0.009667152538895607, + 0.007893526926636696, + 0.005752392578870058, + 0.01347767747938633, + 0.018798690289258957, + 0.00695031275972724, + 0.032724276185035706, + 0.010102025233209133, + 0.00892997719347477, + 0.008619438856840134, + 0.010145272128283978, + 0.013041696511209011, + 0.01461346447467804, + 0.012254697270691395, + 0.01900322735309601, + 0.011260796338319778, + 0.015849318355321884, + 0.012845704331994057, + 0.010986542329192162, + 0.013470658101141453, + 0.013808043673634529, + 0.009682074189186096, + 0.009400810115039349, + 0.014598891139030457, + 0.010598202235996723, + 0.0165227260440588, + 0.011561380699276924, + 0.008259937167167664, + 0.01323605328798294, + 0.009263659827411175, + 0.017868122085928917, + 0.010529473423957825, + 0.0068541583605110645, + 0.011772656813263893, + 0.008230393752455711, + 0.008324391208589077, + 0.01408507488667965, + 0.007884765043854713, + 0.014993955381214619, + 0.00811047200113535, + 0.005122823640704155, + 0.012422816827893257, + 0.006016302853822708, + 0.00866764411330223, + 0.008458063937723637, + 0.010658897459506989, + 0.011259740218520164, + 0.007643002085387707, + 0.005957028828561306, + 0.0056444075889885426, + 0.004533465951681137, + 0.0075990804471075535, + 0.00815628096461296, + 0.005345483310520649, + 0.008963939733803272, + 0.0089131910353899, + 0.005468284245580435, + 0.02035287208855152, + 0.0040642814710736275, + 0.028550341725349426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.016547108069062233 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.018209194764494896 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018209194764494896, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018209194764494896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09770218282938004, + 0.07236295938491821, + 0.07981283962726593, + 0.0773431733250618, + 0.16797903180122375, + 0.15842749178409576, + 0.10645498335361481, + 0.11221533268690109, + 0.0835910439491272, + 0.04411095380783081 + ], + "credit_map": [ + 0.014092013239860535, + 0.009405486285686493, + 0.013999403454363346, + 0.01652107760310173, + 0.0130842300131917, + 0.0055817728862166405, + 0.012378615327179432, + 0.012639586813747883, + 0.015247191302478313, + 0.0053567602299153805, + 0.011184515431523323, + 0.013596303761005402, + 0.006465356331318617, + 0.005930586252361536, + 0.0069088079035282135, + 0.007673438172787428, + 0.009775911457836628, + 0.005836211144924164, + 0.010547373443841934, + 0.012304284609854221, + 0.013407494872808456, + 0.0062430305406451225, + 0.01106442604213953, + 0.010634103789925575, + 0.01134019996970892, + 0.007612964138388634, + 0.011774277314543724, + 0.009353148750960827, + 0.01099466998130083, + 0.011535669676959515, + 0.006242318544536829, + 0.008489931002259254, + 0.023285966366529465, + 0.013945904560387135, + 0.02981177531182766, + 0.013584918342530727, + 0.01663963869214058, + 0.024847595021128654, + 0.021751534193754196, + 0.024111701175570488, + 0.025755256414413452, + 0.013346937485039234, + 0.011119400151073933, + 0.020946020260453224, + 0.01491426583379507, + 0.03755941614508629, + 0.01614065282046795, + 0.018645547330379486, + 0.015386203303933144, + 0.009169470518827438, + 0.0246187224984169, + 0.012295186519622803, + 0.010017147287726402, + 0.01395678985863924, + 0.008574357256293297, + 0.012437105178833008, + 0.021530676633119583, + 0.01256529800593853, + 0.024751437827944756, + 0.01533355563879013, + 0.006314885802567005, + 0.011734926141798496, + 0.008046532049775124, + 0.01193801686167717, + 0.010000419802963734, + 0.013394257053732872, + 0.016150936484336853, + 0.010112925432622433, + 0.00816028006374836, + 0.008813918568193913, + 0.006477202754467726, + 0.010481102392077446, + 0.00490627670660615, + 0.003727502888068557, + 0.005266916006803513, + 0.006637156940996647, + 0.003356877015903592, + 0.005802821833640337, + 0.0030498106498271227, + 0.011363592930138111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.018209194764494896 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.004643728025257587 + ], + "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": [ + 0.07879610359668732, + 0.07343165576457977, + 0.07411428540945053, + 0.10409271717071533, + 0.11112456768751144, + 0.10090382397174835, + 0.13655249774456024, + 0.1381441354751587, + 0.09903991967439651, + 0.0838002860546112 + ], + "credit_map": [ + 0.014873991720378399, + 0.003720578271895647, + 0.014599684625864029, + 0.01277762558311224, + 0.006140889599919319, + 0.0067560793831944466, + 0.00951552763581276, + 0.010411723516881466, + 0.01486292015761137, + 0.0059775966219604015, + 0.009051199071109295, + 0.0125436345115304, + 0.0068733179941773415, + 0.006092339754104614, + 0.010278466157615185, + 0.007752181496471167, + 0.01206399966031313, + 0.00627352437004447, + 0.010400687344372272, + 0.013111358508467674, + 0.010288884863257408, + 0.005866136401891708, + 0.007504458539187908, + 0.008605238050222397, + 0.015229974873363972, + 0.009324775077402592, + 0.021702269092202187, + 0.010868038050830364, + 0.012677975930273533, + 0.01203897874802351, + 0.010689182206988335, + 0.01156152319163084, + 0.014338691718876362, + 0.01009290013462305, + 0.017050817608833313, + 0.010020109824836254, + 0.01676277443766594, + 0.016952255740761757, + 0.011510653421282768, + 0.01439636293798685, + 0.019339807331562042, + 0.007087770849466324, + 0.008716232143342495, + 0.019305208697915077, + 0.005779200233519077, + 0.016385801136493683, + 0.007706375326961279, + 0.016583431512117386, + 0.023016301915049553, + 0.00855996459722519, + 0.041925206780433655, + 0.015065825544297695, + 0.007706568576395512, + 0.018156874924898148, + 0.006626997608691454, + 0.015494759194552898, + 0.02311176247894764, + 0.012337018735706806, + 0.022469472140073776, + 0.016465958207845688, + 0.010491360910236835, + 0.02318008430302143, + 0.013293847441673279, + 0.016794627532362938, + 0.01368001289665699, + 0.014628727920353413, + 0.019767777994275093, + 0.012132076546549797, + 0.007798280566930771, + 0.01011376827955246, + 0.00846901349723339, + 0.012450264766812325, + 0.00974216591566801, + 0.009239443577826023, + 0.009742002934217453, + 0.012267326936125755, + 0.006186119746416807, + 0.013186934404075146, + 0.00401287991553545, + 0.019423414021730423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.004643728025257587 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.06205030158162117 + ], + "opposite_logprob_tokens": [ + -2.8120503425598145 + ], + "logprob_reward": -0.06205030158162117, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205030158162117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06222165375947952, + 0.056695833802223206, + 0.05041592940688133, + 0.0893537849187851, + 0.11377736181020737, + 0.1822553277015686, + 0.20283184945583344, + 0.0766897201538086, + 0.08415867388248444, + 0.08159986883401871 + ], + "credit_map": [ + 0.01076036412268877, + 0.0058768559247255325, + 0.00878205243498087, + 0.009801867417991161, + 0.005156894214451313, + 0.004178646486252546, + 0.007159657776355743, + 0.010505310259759426, + 0.009427236393094063, + 0.004591119475662708, + 0.008265916258096695, + 0.010662254877388477, + 0.005977442488074303, + 0.004240969195961952, + 0.007049693260341883, + 0.0064811985939741135, + 0.006142063532024622, + 0.0040835668332874775, + 0.006460551172494888, + 0.011402676813304424, + 0.00435595540329814, + 0.003543150844052434, + 0.00600086385384202, + 0.008427094668149948, + 0.009394216351211071, + 0.005565640516579151, + 0.018346630036830902, + 0.006588486954569817, + 0.008092338219285011, + 0.008216332644224167, + 0.008352786302566528, + 0.02479735016822815, + 0.016606425866484642, + 0.007809044327586889, + 0.02319897525012493, + 0.009093377739191055, + 0.014151286333799362, + 0.01960924081504345, + 0.012707249261438847, + 0.010601756162941456, + 0.0288714412599802, + 0.011952150613069534, + 0.025463279336690903, + 0.018615582957863808, + 0.013673210516571999, + 0.03240910917520523, + 0.02019086666405201, + 0.03107968159019947, + 0.03496036306023598, + 0.013950064778327942, + 0.07100722938776016, + 0.026819955557584763, + 0.015400618314743042, + 0.017706023529171944, + 0.008808876387774944, + 0.014178714714944363, + 0.011399954557418823, + 0.008819878101348877, + 0.012065019458532333, + 0.00748170493170619, + 0.006755673326551914, + 0.013474722392857075, + 0.007514893542975187, + 0.009177865460515022, + 0.01428148988634348, + 0.014344891533255577, + 0.014077865518629551, + 0.008944906294345856, + 0.008715638890862465, + 0.007376465480774641, + 0.006193250883370638, + 0.010224160738289356, + 0.00794936716556549, + 0.004615664016455412, + 0.0072550224140286446, + 0.007449704688042402, + 0.005917613860219717, + 0.02493966370820999, + 0.004201123956590891, + 0.01927170343697071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.06205030158162117 + }, + { + "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 initial shot a macro close-up focusing on an ear covered in dense, light-brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, golden needle with a metallic sheen visible being pulled out of the ear canal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back to reveal a wider scene as the needle exits the ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the needle transform into a staff with heavy golden hoops and a deep ebony hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a dazzling golden light erupting from the needle against a dark 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": [ + -0.043063629418611526 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.043063629418611526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063629418611526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05305621773004532, + 0.07406622171401978, + 0.084947869181633, + 0.04373195394873619, + 0.09082119166851044, + 0.048837751150131226, + 0.04537087678909302, + 0.316819965839386, + 0.1261827051639557, + 0.11616522073745728 + ], + "credit_map": [ + 0.008025755174458027, + 0.004608083050698042, + 0.008012196980416775, + 0.0070001957938075066, + 0.005842314567416906, + 0.004564396571367979, + 0.007367593236267567, + 0.007635683752596378, + 0.017366450279951096, + 0.004153681453317404, + 0.014213352464139462, + 0.010158761404454708, + 0.005228658206760883, + 0.005912627559155226, + 0.0075515261851251125, + 0.009481173940002918, + 0.0237994734197855, + 0.003767350921407342, + 0.017921730875968933, + 0.010717381723225117, + 0.005191633477807045, + 0.006614317651838064, + 0.005992427002638578, + 0.010943559929728508, + 0.0056082699447870255, + 0.0030501489527523518, + 0.008339064195752144, + 0.005347040016204119, + 0.004396809730678797, + 0.00409338716417551, + 0.004443855956196785, + 0.008453380316495895, + 0.014715813100337982, + 0.006906419061124325, + 0.013315069489181042, + 0.010246891528367996, + 0.007819289341568947, + 0.012952200137078762, + 0.011831407435238361, + 0.01303410530090332, + 0.007933155633509159, + 0.0031606543343514204, + 0.00603088503703475, + 0.0072451806627213955, + 0.004018691368401051, + 0.010603590868413448, + 0.00544756418094039, + 0.004398035351186991, + 0.007023198995739222, + 0.005286188330501318, + 0.01181798242032528, + 0.00529251666739583, + 0.004345477093011141, + 0.0043877605348825455, + 0.0035773811396211386, + 0.0036403737030923367, + 0.06397977471351624, + 0.040740154683589935, + 0.07649876922369003, + 0.04514710605144501, + 0.012979992665350437, + 0.026452887803316116, + 0.020729705691337585, + 0.03029158152639866, + 0.021859386935830116, + 0.020579615607857704, + 0.02331014722585678, + 0.01627035066485405, + 0.009544131346046925, + 0.010308980010449886, + 0.008783395402133465, + 0.015526705421507359, + 0.0069565074518322945, + 0.006785064935684204, + 0.008547559380531311, + 0.030189858749508858, + 0.00702632637694478, + 0.013567546382546425, + 0.004820540081709623, + 0.03827183321118355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.043063629418611526 + } + ], + "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/0000167/group_0/index_tensor([6713, 6713])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_tensor([6713, 6713])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7dc5ff392b3d2409bb904e4e189ab8e01ca6924a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_tensor([6713, 6713])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme low-angle shot, looking upward from the ground, the camera tilts steadily upward, closely following the foot of a massive mechanical armor. This foot is composed of battle-worn, dark gray metal armor, its surface covered in scratches and intricate wiring. A massive thruster nozzle at the center of its sole suddenly unleashes a stream of blindingly hot blue-white plasma. This destructive energy instantly pulverizes the concrete below, unleashing a torrent of dense smoke and rubble of varying sizes, which scatters in all directions in a powerful shockwave. Driven by immense force, the behemoth rises vertically in a steady and unstoppable manner. The blazing air creates distinct ripples in the air, casting a dazzling light on the swirling cloud of smoke and dust and the armor's rough metal legs, creating a violent flicker of light and shadow. The entire picture is filled with overwhelming power and apocalyptic grandeur, with tones dominated by cold gray and metallic colors. Only the dazzling blue and white of the energy beam becomes the only color core in the entire scene, full of destructive vitality.", + "seed": 23219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_6713_seed_23219_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": -0.018726359214633705, + "vlm_logprob": -0.018726359214633705, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.018726359214633705, + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03389321640133858 + ], + "opposite_logprob_tokens": [ + -3.53389310836792 + ], + "logprob_reward": -0.03389321640133858, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03389321640133858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3071429431438446, + 0.04164445400238037, + 0.09195957332849503, + 0.18757013976573944, + 0.09605573117733002, + 0.02472037263214588, + 0.01592419669032097, + 0.031500283628702164, + 0.02924082987010479, + 0.17424151301383972 + ], + "credit_map": [ + 0.09435094147920609, + 0.012799861840903759, + 0.013421045616269112, + 0.011764184571802616, + 0.013858404941856861, + 0.023981695994734764, + 0.026288682594895363, + 0.11067812889814377, + 0.004737402778118849, + 0.005481199361383915, + 0.003673631465062499, + 0.004116042051464319, + 0.006483670324087143, + 0.005028358194977045, + 0.006059038918465376, + 0.00606510927900672, + 0.012294345535337925, + 0.007742586079984903, + 0.01585925929248333, + 0.013143674470484257, + 0.007294792216271162, + 0.013191337697207928, + 0.011589919216930866, + 0.01084365788847208, + 0.012695690616965294, + 0.013477620668709278, + 0.02904844842851162, + 0.019581209868192673, + 0.01165456511080265, + 0.013029040768742561, + 0.05629701167345047, + 0.03178653493523598, + 0.006190189626067877, + 0.006404168903827667, + 0.02296900935471058, + 0.007948161102831364, + 0.003537385957315564, + 0.012746627442538738, + 0.02903790771961212, + 0.007222277577966452, + 0.0028889684472233057, + 0.002720975549891591, + 0.003471142379567027, + 0.00260060909204185, + 0.0026927525177598, + 0.003380104433745146, + 0.00230445247143507, + 0.0046613686718046665, + 0.001476919511333108, + 0.0018070490332320333, + 0.002125660190358758, + 0.0034487054217606783, + 0.0009190019336529076, + 0.0011341614881530404, + 0.0027379102539271116, + 0.002274787751957774, + 0.002969409106299281, + 0.0038877942133694887, + 0.006403644103556871, + 0.003653888590633869, + 0.002977276686578989, + 0.0028086421079933643, + 0.004324047826230526, + 0.004475582856684923, + 0.002018690574914217, + 0.0031982650980353355, + 0.00392580172047019, + 0.0049773696810007095, + 0.004164157900959253, + 0.0028598043136298656, + 0.004582055378705263, + 0.003514684271067381, + 0.013493637554347515, + 0.019351724535226822, + 0.015199503861367702, + 0.01998736895620823, + 0.023172495886683464, + 0.014494777657091618, + 0.034539442509412766, + 0.03400255739688873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.03389321640133858 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006756437011063099 + ], + "opposite_logprob_tokens": [ + -5.00675630569458 + ], + "logprob_reward": -0.006756437011063099, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006756437011063099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1836499273777008, + 0.06204455718398094, + 0.04136814922094345, + 0.28340697288513184, + 0.09312986582517624, + 0.045814357697963715, + 0.11461939662694931, + 0.04134207218885422, + 0.028440484777092934, + 0.10618416219949722 + ], + "credit_map": [ + 0.05104547366499901, + 0.008896267972886562, + 0.008976511657238007, + 0.011723055504262447, + 0.008767526596784592, + 0.018393486738204956, + 0.015710841864347458, + 0.06013677641749382, + 0.006410158239305019, + 0.007462369278073311, + 0.009301730431616306, + 0.008962004445493221, + 0.007015462499111891, + 0.006644286680966616, + 0.009112868458032608, + 0.007135681342333555, + 0.006085237953811884, + 0.003490051720291376, + 0.008978331461548805, + 0.005760044325143099, + 0.0034002880565822124, + 0.004077700432389975, + 0.004610329866409302, + 0.0049661691300570965, + 0.01500033587217331, + 0.01251170877367258, + 0.09778895229101181, + 0.030905254185199738, + 0.011022905819118023, + 0.01114910189062357, + 0.08317331969738007, + 0.021855419501662254, + 0.005786416120827198, + 0.004016191698610783, + 0.015502788126468658, + 0.007550232112407684, + 0.0032648092601448298, + 0.024430496618151665, + 0.02219814993441105, + 0.010380790568888187, + 0.007167167495936155, + 0.004257828928530216, + 0.008876877836883068, + 0.004932827781885862, + 0.0040135676972568035, + 0.00538990180939436, + 0.004020116291940212, + 0.007156074978411198, + 0.00892092939466238, + 0.013322005979716778, + 0.02401578612625599, + 0.026398099958896637, + 0.005562556441873312, + 0.007644035387784243, + 0.012721912004053593, + 0.016034070402383804, + 0.0034057132434099913, + 0.0054002185352146626, + 0.008464050479233265, + 0.005678554996848106, + 0.004277687985450029, + 0.005321302451193333, + 0.0046291532926261425, + 0.004165394231677055, + 0.0029006663244217634, + 0.003569426480680704, + 0.0027324403636157513, + 0.004743053112179041, + 0.004876457154750824, + 0.0026903243269771338, + 0.004685531370341778, + 0.0022425863426178694, + 0.012803270481526852, + 0.011338422074913979, + 0.010816679336130619, + 0.018313370645046234, + 0.013626480475068092, + 0.008665056899189949, + 0.017127864062786102, + 0.013493028469383717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.006756437011063099 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48691317439079285, + 0.06604581326246262, + 0.047467708587646484, + 0.20001043379306793, + 0.06669024378061295, + 0.03766361624002457, + 0.02252848632633686, + 0.026386933401226997, + 0.026564469560980797, + 0.0197291262447834 + ], + "credit_map": [ + 0.1487700194120407, + 0.017260150983929634, + 0.022136574611067772, + 0.016641026362776756, + 0.02229098416864872, + 0.040236618369817734, + 0.04352231323719025, + 0.17605547606945038, + 0.008441157639026642, + 0.005970984231680632, + 0.011393185704946518, + 0.007421106565743685, + 0.0069001540541648865, + 0.006862280424684286, + 0.010363670997321606, + 0.008693274110555649, + 0.007342181634157896, + 0.003019841853529215, + 0.00988661590963602, + 0.005631140433251858, + 0.0047476449981331825, + 0.005376971792429686, + 0.005023520905524492, + 0.00643979012966156, + 0.010264243930578232, + 0.00679116602987051, + 0.0708538144826889, + 0.020368115976452827, + 0.006261948496103287, + 0.0060120620764791965, + 0.06238806992769241, + 0.017071010544896126, + 0.0032926828134804964, + 0.003861659672111273, + 0.00591447576880455, + 0.004288724157959223, + 0.003222472034394741, + 0.006982993334531784, + 0.02521883137524128, + 0.013908407650887966, + 0.005310249049216509, + 0.003740498796105385, + 0.009251730516552925, + 0.004865999333560467, + 0.003394266590476036, + 0.0032286737114191055, + 0.003664518240839243, + 0.0042076800018548965, + 0.0020456071943044662, + 0.002946378430351615, + 0.002163337543606758, + 0.005537079647183418, + 0.0012816819362342358, + 0.0025873766280710697, + 0.0030028540641069412, + 0.002964172512292862, + 0.001778972102329135, + 0.0030566949862986803, + 0.005375623237341642, + 0.0035243690945208073, + 0.0030751884914934635, + 0.0034565599635243416, + 0.0031061226036399603, + 0.0030134045518934727, + 0.002319600898772478, + 0.002903346437960863, + 0.0038653460796922445, + 0.004095504060387611, + 0.0034068364184349775, + 0.0021688928827643394, + 0.0049062855541706085, + 0.0028986588586121798, + 0.002131236484274268, + 0.0020775864832103252, + 0.0020839241333305836, + 0.0030290333088487387, + 0.002469224389642477, + 0.0015108896186575294, + 0.0032563842833042145, + 0.003170848125591874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03370778262615204 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734768394380808 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734768394380808, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734768394380808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3185572028160095, + 0.036178845912218094, + 0.03147909417748451, + 0.2712596654891968, + 0.13706143200397491, + 0.0494503527879715, + 0.0450059212744236, + 0.03920915722846985, + 0.053589753806591034, + 0.018208570778369904 + ], + "credit_map": [ + 0.09849399328231812, + 0.010670928284525871, + 0.014313917607069016, + 0.009136753156781197, + 0.014226003549993038, + 0.02494228072464466, + 0.028210166841745377, + 0.11856315284967422, + 0.0038140600081533194, + 0.004102187231183052, + 0.0050958204083144665, + 0.006658964790403843, + 0.003870698856189847, + 0.002811684273183346, + 0.005684395786374807, + 0.004141033627092838, + 0.004341212101280689, + 0.002318882616236806, + 0.004991989582777023, + 0.004582012537866831, + 0.003449388314038515, + 0.003497842000797391, + 0.003416763385757804, + 0.0048810020089149475, + 0.013667888939380646, + 0.009369530715048313, + 0.09709499776363373, + 0.02805144712328911, + 0.008764921687543392, + 0.009142475202679634, + 0.0835866779088974, + 0.02158174104988575, + 0.009107932448387146, + 0.0038218414410948753, + 0.017997438088059425, + 0.010323634371161461, + 0.005711854435503483, + 0.04248834773898125, + 0.03547808155417442, + 0.012132301926612854, + 0.005619129166007042, + 0.005777288693934679, + 0.005457248073071241, + 0.0049937753938138485, + 0.006369142793118954, + 0.0063093844801187515, + 0.00932689942419529, + 0.005597483366727829, + 0.0035430488642305136, + 0.005473938304930925, + 0.0048331026919186115, + 0.01222141832113266, + 0.002276772167533636, + 0.004090622067451477, + 0.006235513836145401, + 0.006331506185233593, + 0.0025986710097640753, + 0.00752631202340126, + 0.0052874465472996235, + 0.005134828854352236, + 0.005759092513471842, + 0.003891957923769951, + 0.004177879076451063, + 0.004832970444113016, + 0.004209600854665041, + 0.0046391561627388, + 0.006175522226840258, + 0.007479299791157246, + 0.009872612543404102, + 0.0043715545907616615, + 0.009949536062777042, + 0.006892472971230745, + 0.001542158075608313, + 0.0019415321294218302, + 0.0019531957805156708, + 0.002330512274056673, + 0.002440745010972023, + 0.0015338017838075757, + 0.0035386832896620035, + 0.002927943132817745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.006734768394380808 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539591640233994 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539591640233994, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539591640233994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4481760859489441, + 0.07252878695726395, + 0.06429439038038254, + 0.08655440807342529, + 0.14080725610256195, + 0.03198130428791046, + 0.017775218933820724, + 0.06452038884162903, + 0.028720363974571228, + 0.044641777873039246 + ], + "credit_map": [ + 0.14174145460128784, + 0.016268840059638023, + 0.019047405570745468, + 0.013905523344874382, + 0.02066025882959366, + 0.03184930607676506, + 0.03897349163889885, + 0.165729820728302, + 0.00752599909901619, + 0.00880796555429697, + 0.01096486859023571, + 0.012227597646415234, + 0.007495079655200243, + 0.006661484483629465, + 0.009719781577587128, + 0.009126007556915283, + 0.008122744038701057, + 0.004859777167439461, + 0.009079894050955772, + 0.007789100520312786, + 0.007662932854145765, + 0.009846177883446217, + 0.007787043694406748, + 0.00914672203361988, + 0.007225855719298124, + 0.004219863563776016, + 0.012060343287885189, + 0.007157895248383284, + 0.004471172578632832, + 0.005804962012916803, + 0.028448650613427162, + 0.017165670171380043, + 0.011557278223335743, + 0.007600039709359407, + 0.0303195733577013, + 0.01404639519751072, + 0.006457463372498751, + 0.046906132251024246, + 0.011561520397663116, + 0.01235885452479124, + 0.004291230347007513, + 0.0036129229702055454, + 0.005717541091144085, + 0.005327859427779913, + 0.003447530558332801, + 0.003379307920113206, + 0.002874498488381505, + 0.0033304120879620314, + 0.0015728892758488655, + 0.0021677010226994753, + 0.0022826583590358496, + 0.003137716557830572, + 0.001032702624797821, + 0.0015496305422857404, + 0.003641823772341013, + 0.0023900968953967094, + 0.005525125190615654, + 0.011724217794835567, + 0.008677573874592781, + 0.00947781652212143, + 0.0066968039609491825, + 0.005645806901156902, + 0.008021816611289978, + 0.00875123031437397, + 0.0026724711060523987, + 0.003283649915829301, + 0.0033967571798712015, + 0.004587531089782715, + 0.004231833852827549, + 0.0023455952759832144, + 0.004564540460705757, + 0.003637984162196517, + 0.004115650430321693, + 0.005099826026707888, + 0.004837608430534601, + 0.006484366487711668, + 0.005124720279127359, + 0.0037087472155690193, + 0.007600311655551195, + 0.007670546881854534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.012539591640233994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, looking upward from the ground, the camera tilts steadily upward, closely following the foot of a massive mechanical armor. This foot is composed of battle-worn, dark gray metal armor, its surface covered in scratches and intricate wiring. A massive thruster nozzle at the center of its sole suddenly unleashes a stream of blindingly hot blue-white plasma. This destructive energy instantly pulverizes the concrete below, unleashing a torrent of dense smoke and rubble of varying sizes, which scatters in all directions in a powerful shockwave. Driven by immense force, the behemoth rises vertically in a steady and unstoppable manner. The blazing air creates distinct ripples in the air, casting a dazzling light on the swirling cloud of smoke and dust and the armor's rough metal legs, creating a violent flicker of light and shadow. The entire picture is filled with overwhelming power and apocalyptic grandeur, with tones dominated by cold gray and metallic colors. Only the dazzling blue and white of the energy beam becomes the only color core in the entire scene, full of destructive vitality.", + "seed": 23220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_6713_seed_23220_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": -0.031724195927381516, + "vlm_logprob": -0.031724195927381516, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.031724195927381516, + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07660295069217682 + ], + "opposite_logprob_tokens": [ + -2.7016029357910156 + ], + "logprob_reward": -0.07660295069217682, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07660295069217682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12592296302318573, + 0.22635135054588318, + 0.06422656774520874, + 0.09051376581192017, + 0.060813385993242264, + 0.023575685918331146, + 0.09210162609815598, + 0.06594621390104294, + 0.16848452389240265, + 0.08206387609243393 + ], + "credit_map": [ + 0.010937602259218693, + 0.01600389927625656, + 0.01572227105498314, + 0.02036074362695217, + 0.012789027765393257, + 0.009043421596288681, + 0.02442280948162079, + 0.016643183305859566, + 0.01760142296552658, + 0.0859292596578598, + 0.011096318252384663, + 0.04277230054140091, + 0.01700018160045147, + 0.015344777144491673, + 0.017876382917165756, + 0.01873070001602173, + 0.008296502754092216, + 0.0075070858001708984, + 0.011803650297224522, + 0.009053532965481281, + 0.004864381160587072, + 0.009104990400373936, + 0.007868068292737007, + 0.005728358402848244, + 0.008557295426726341, + 0.006825751159340143, + 0.00783536396920681, + 0.0163214523345232, + 0.008955015800893307, + 0.009110691025853157, + 0.025721047073602676, + 0.007187152747064829, + 0.005740154068917036, + 0.00816970132291317, + 0.0071901557967066765, + 0.005063317716121674, + 0.013377973809838295, + 0.009269597940146923, + 0.00490531325340271, + 0.007097168359905481, + 0.0029802098870277405, + 0.0035716877318918705, + 0.0028313344810158014, + 0.00284374481998384, + 0.004721693228930235, + 0.002392987720668316, + 0.0022865482605993748, + 0.0019474810687825084, + 0.011745188385248184, + 0.008619727566838264, + 0.013225354254245758, + 0.01322917826473713, + 0.011127864941954613, + 0.011869439855217934, + 0.011754804290831089, + 0.010530072264373302, + 0.0063212234526872635, + 0.014011700637638569, + 0.005829099100083113, + 0.007625121623277664, + 0.00876524392515421, + 0.006923535373061895, + 0.00696443160995841, + 0.009505854919552803, + 0.005735896527767181, + 0.004321447107940912, + 0.007010552566498518, + 0.00329161761328578, + 0.06533388048410416, + 0.06904465705156326, + 0.0074400343000888824, + 0.006306433118879795, + 0.008885964751243591, + 0.009763705544173717, + 0.011737010441720486, + 0.011665252968668938, + 0.01446938794106245, + 0.00838865339756012, + 0.009134863503277302, + 0.008019035682082176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.07660295069217682 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029824471101164818 + ], + "opposite_logprob_tokens": [ + -3.5298244953155518 + ], + "logprob_reward": -0.029824471101164818, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029824471101164818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1071392148733139, + 0.09226120263338089, + 0.05103786289691925, + 0.17609523236751556, + 0.05405336245894432, + 0.03380493447184563, + 0.09824024885892868, + 0.06339292228221893, + 0.23444883525371552, + 0.0895262211561203 + ], + "credit_map": [ + 0.010331501252949238, + 0.012502005323767662, + 0.018008964136242867, + 0.0133356973528862, + 0.01145574264228344, + 0.007108554244041443, + 0.016380498185753822, + 0.018016250804066658, + 0.011199425905942917, + 0.01069918368011713, + 0.008482138626277447, + 0.015704911202192307, + 0.010187126696109772, + 0.01232451293617487, + 0.011185361072421074, + 0.012478540651500225, + 0.006629867944866419, + 0.0071069528348743916, + 0.005769048351794481, + 0.007740050554275513, + 0.005767147522419691, + 0.009561415761709213, + 0.004741317592561245, + 0.00372205744497478, + 0.011594411917030811, + 0.01231706328690052, + 0.006145692430436611, + 0.06397182494401932, + 0.008881263434886932, + 0.0073472922667860985, + 0.05560851842164993, + 0.010229173116385937, + 0.004306507296860218, + 0.006366918329149485, + 0.010084395296871662, + 0.00577315641567111, + 0.008440379053354263, + 0.008198605850338936, + 0.004465167876332998, + 0.006418230943381786, + 0.005429608281701803, + 0.004408263601362705, + 0.005195023957639933, + 0.003926307428628206, + 0.005491794552654028, + 0.0035719608422368765, + 0.0027049852069467306, + 0.0030769873410463333, + 0.011851438321173191, + 0.008196935057640076, + 0.009864624589681625, + 0.015510952100157738, + 0.010893577709794044, + 0.012278713285923004, + 0.017259223386645317, + 0.012384772300720215, + 0.005725527182221413, + 0.009305810555815697, + 0.00577740091830492, + 0.01175254862755537, + 0.005370966158807278, + 0.007409896235913038, + 0.008189551532268524, + 0.009861217811703682, + 0.0116726104170084, + 0.008090758696198463, + 0.013740358874201775, + 0.007157331332564354, + 0.0815148651599884, + 0.09160780161619186, + 0.012509249150753021, + 0.008155846036970615, + 0.008903945796191692, + 0.010080117732286453, + 0.012066952884197235, + 0.01292730774730444, + 0.014585149474442005, + 0.010163674131035805, + 0.011853578500449657, + 0.008945486508309841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.029824471101164818 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358619332313538 + ], + "opposite_logprob_tokens": [ + -3.6513586044311523 + ], + "logprob_reward": -0.026358619332313538, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358619332313538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15028692781925201, + 0.10928193479776382, + 0.09630845487117767, + 0.07924909889698029, + 0.07228557765483856, + 0.07298571616411209, + 0.11772774904966354, + 0.034179892390966415, + 0.20616905391216278, + 0.061525553464889526 + ], + "credit_map": [ + 0.016989801079034805, + 0.02014772780239582, + 0.020254530012607574, + 0.019820652902126312, + 0.01769413985311985, + 0.011083223856985569, + 0.022851968184113503, + 0.02144489251077175, + 0.01478973962366581, + 0.016186196357011795, + 0.01235196739435196, + 0.008059716783463955, + 0.015076729469001293, + 0.008140252903103828, + 0.013570170849561691, + 0.021107163280248642, + 0.009909318760037422, + 0.014161682687699795, + 0.02338455244898796, + 0.012220620177686214, + 0.007794511504471302, + 0.012569429352879524, + 0.0069845328107476234, + 0.009283807128667831, + 0.01221481617540121, + 0.010823079384863377, + 0.007453443948179483, + 0.015232968144118786, + 0.0058500077575445175, + 0.010731521993875504, + 0.007747061084955931, + 0.009196207858622074, + 0.008698588237166405, + 0.009068266488611698, + 0.01240471564233303, + 0.007591151632368565, + 0.013515390455722809, + 0.0081521887332201, + 0.005412975791841745, + 0.0074423011392354965, + 0.007197220344096422, + 0.008807371370494366, + 0.011256196536123753, + 0.009478392079472542, + 0.010953795164823532, + 0.010874857194721699, + 0.0066771735437214375, + 0.007740713655948639, + 0.011525128968060017, + 0.012163805775344372, + 0.012119501829147339, + 0.0128531688824296, + 0.017952701076865196, + 0.018696002662181854, + 0.021658383309841156, + 0.010759063996374607, + 0.004244592972099781, + 0.005919444374740124, + 0.0040515814907848835, + 0.004100847523659468, + 0.0037741551641374826, + 0.003372695529833436, + 0.004321974702179432, + 0.004394602496176958, + 0.005983064882457256, + 0.006495490204542875, + 0.008271381258964539, + 0.00448249327018857, + 0.08009053766727448, + 0.08505630493164062, + 0.00888997782021761, + 0.006899826694279909, + 0.007042593788355589, + 0.005617690738290548, + 0.007738233543932438, + 0.010654976591467857, + 0.010546112433075905, + 0.006714168470352888, + 0.008071503601968288, + 0.005140282213687897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.026358619332313538 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254032090306282 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254032090306282, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254032090306282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10511321574449539, + 0.04049066826701164, + 0.045962970703840256, + 0.07358826696872711, + 0.08504955470561981, + 0.1226632222533226, + 0.12312132865190506, + 0.09736232459545135, + 0.1143454760313034, + 0.19230300188064575 + ], + "credit_map": [ + 0.011518876068294048, + 0.012034530751407146, + 0.013723941519856453, + 0.013929278589785099, + 0.012153568677604198, + 0.006248454563319683, + 0.01932079903781414, + 0.0161837637424469, + 0.004117538221180439, + 0.010389959439635277, + 0.0028958474285900593, + 0.004803773481398821, + 0.003935509826987982, + 0.005527866072952747, + 0.005267797503620386, + 0.0035523739643394947, + 0.005522886756807566, + 0.005900092888623476, + 0.010071092285215855, + 0.005854330491274595, + 0.003649357473477721, + 0.0051895566284656525, + 0.00653143459931016, + 0.0032442216761410236, + 0.006787785328924656, + 0.005992360413074493, + 0.0034356468822807074, + 0.022599149495363235, + 0.004612588789314032, + 0.004498621448874474, + 0.021521490067243576, + 0.004140630830079317, + 0.007849435321986675, + 0.011899313889443874, + 0.009621957316994667, + 0.013991904444992542, + 0.015426702797412872, + 0.00911056250333786, + 0.00770212896168232, + 0.009447549469769001, + 0.014693360775709152, + 0.014683572575449944, + 0.020209552720189095, + 0.013210213743150234, + 0.02016109973192215, + 0.016174690797924995, + 0.011894898489117622, + 0.011635832488536835, + 0.013709410093724728, + 0.011957738548517227, + 0.01672687940299511, + 0.012977736070752144, + 0.017506541684269905, + 0.015671372413635254, + 0.02369808219373226, + 0.010873567312955856, + 0.009144539944827557, + 0.014113721437752247, + 0.009876996278762817, + 0.014242851175367832, + 0.0075632501393556595, + 0.008283968083560467, + 0.015113545581698418, + 0.019023455679416656, + 0.009270288981497288, + 0.012530118227005005, + 0.014873171225190163, + 0.007931843400001526, + 0.024954812601208687, + 0.02009693905711174, + 0.01257083099335432, + 0.012117470614612103, + 0.01877923496067524, + 0.027771111577749252, + 0.024659547954797745, + 0.03767802566289902, + 0.020574215799570084, + 0.016436101868748665, + 0.028811823576688766, + 0.017592953518033028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.005254032090306282 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058090642094612 + ], + "opposite_logprob_tokens": [ + -3.895581007003784 + ], + "logprob_reward": -0.02058090642094612, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058090642094612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15657955408096313, + 0.04667613282799721, + 0.026740124449133873, + 0.11440471559762955, + 0.13676190376281738, + 0.037734393030405045, + 0.14828790724277496, + 0.10536542534828186, + 0.08713225275278091, + 0.14031760394573212 + ], + "credit_map": [ + 0.014540673233568668, + 0.01698971912264824, + 0.016211654990911484, + 0.02402515895664692, + 0.014382174238562584, + 0.013904483057558537, + 0.02802198939025402, + 0.02850370667874813, + 0.006216774228960276, + 0.005939452443271875, + 0.0041875336319208145, + 0.005499729420989752, + 0.0064185201190412045, + 0.005838205572217703, + 0.006158826407045126, + 0.006417092867195606, + 0.0031864505726844072, + 0.003529744688421488, + 0.004210552666336298, + 0.003233115654438734, + 0.002703191479668021, + 0.004802957642823458, + 0.0025406922213733196, + 0.002533417195081711, + 0.009176590479910374, + 0.008633296936750412, + 0.005052610766142607, + 0.033114392310380936, + 0.007100362330675125, + 0.008608495816588402, + 0.03516389802098274, + 0.0075550624169409275, + 0.01053052768111229, + 0.016750555485486984, + 0.019492385908961296, + 0.012510017491877079, + 0.03009314462542534, + 0.021040115505456924, + 0.010033423081040382, + 0.016311733052134514, + 0.005043605342507362, + 0.004475017543882132, + 0.007690057158470154, + 0.00397156598046422, + 0.005870831198990345, + 0.004547053948044777, + 0.003752944990992546, + 0.0023833168670535088, + 0.017883026972413063, + 0.013927836902439594, + 0.02017831616103649, + 0.024871421977877617, + 0.017367009073495865, + 0.018267856910824776, + 0.01877943053841591, + 0.01701301522552967, + 0.009045197628438473, + 0.015414831228554249, + 0.0069091953337192535, + 0.014304989017546177, + 0.010988350957632065, + 0.008309838362038136, + 0.02454632893204689, + 0.01584669016301632, + 0.005967460572719574, + 0.007437916472554207, + 0.01058001909404993, + 0.006395782809704542, + 0.022179067134857178, + 0.022275447845458984, + 0.006634809076786041, + 0.005661750677973032, + 0.015450281091034412, + 0.0152742275968194, + 0.024009589105844498, + 0.018340155482292175, + 0.024273863062262535, + 0.015196546912193298, + 0.016317110508680344, + 0.011455836705863476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.02058090642094612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, looking upward from the ground, the camera tilts steadily upward, closely following the foot of a massive mechanical armor. This foot is composed of battle-worn, dark gray metal armor, its surface covered in scratches and intricate wiring. A massive thruster nozzle at the center of its sole suddenly unleashes a stream of blindingly hot blue-white plasma. This destructive energy instantly pulverizes the concrete below, unleashing a torrent of dense smoke and rubble of varying sizes, which scatters in all directions in a powerful shockwave. Driven by immense force, the behemoth rises vertically in a steady and unstoppable manner. The blazing air creates distinct ripples in the air, casting a dazzling light on the swirling cloud of smoke and dust and the armor's rough metal legs, creating a violent flicker of light and shadow. The entire picture is filled with overwhelming power and apocalyptic grandeur, with tones dominated by cold gray and metallic colors. Only the dazzling blue and white of the energy beam becomes the only color core in the entire scene, full of destructive vitality.", + "seed": 23221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_6713_seed_23221_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": -0.04235549494624138, + "vlm_logprob": -0.04235549494624138, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04235549494624138, + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.107691690325737 + ], + "opposite_logprob_tokens": [ + -2.357691764831543 + ], + "logprob_reward": -0.107691690325737, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.107691690325737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15140163898468018, + 0.05045321211218834, + 0.250240683555603, + 0.04288849979639053, + 0.11998055875301361, + 0.03165075182914734, + 0.09900771826505661, + 0.05403904244303703, + 0.05564303323626518, + 0.14469484984874725 + ], + "credit_map": [ + 0.011427467688918114, + 0.011539699509739876, + 0.015240605920553207, + 0.020639406517148018, + 0.016801759600639343, + 0.018018027767539024, + 0.03140577673912048, + 0.02632889337837696, + 0.0059700836427509785, + 0.004007060546427965, + 0.0066177756525576115, + 0.005201342981308699, + 0.0077896639704704285, + 0.007332265377044678, + 0.006778254639357328, + 0.0067567662335932255, + 0.008504019118845463, + 0.0058098952285945415, + 0.09245169162750244, + 0.01517816074192524, + 0.007327997591346502, + 0.011994021944701672, + 0.013862254098057747, + 0.0951126366853714, + 0.004684263374656439, + 0.004675125703215599, + 0.0051513006910681725, + 0.006556330714374781, + 0.0041882190853357315, + 0.0038600927218794823, + 0.007220735773444176, + 0.006552430335432291, + 0.008011894300580025, + 0.007400468923151493, + 0.026067879050970078, + 0.03037656843662262, + 0.010032527148723602, + 0.01287058088928461, + 0.010506322607398033, + 0.014714314602315426, + 0.004751278553158045, + 0.002971445443108678, + 0.003966871649026871, + 0.006048134062439203, + 0.002841019770130515, + 0.0031489618122577667, + 0.004181642550975084, + 0.0037414005491882563, + 0.011628275737166405, + 0.011244979687035084, + 0.013856726698577404, + 0.014692549593746662, + 0.013796314597129822, + 0.010149974375963211, + 0.013830689713358879, + 0.009808208793401718, + 0.004784876946359873, + 0.006855425890535116, + 0.005217595957219601, + 0.009179889224469662, + 0.007477927953004837, + 0.005915991496294737, + 0.008929704315960407, + 0.005677629727870226, + 0.004642795771360397, + 0.007577158510684967, + 0.005764234345406294, + 0.006712204311043024, + 0.009536553174257278, + 0.005313182715326548, + 0.008468803949654102, + 0.007628096267580986, + 0.018911436200141907, + 0.012809227220714092, + 0.025980474427342415, + 0.01792878657579422, + 0.012582745403051376, + 0.01835240237414837, + 0.01963614486157894, + 0.018493637442588806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.107691690325737 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813860937952995 + ], + "opposite_logprob_tokens": [ + -3.2881386280059814 + ], + "logprob_reward": -0.03813860937952995, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813860937952995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044160258024930954, + 0.0539017952978611, + 0.457589715719223, + 0.03492835909128189, + 0.03872667998075485, + 0.041638415306806564, + 0.08780508488416672, + 0.07511495053768158, + 0.08949694782495499, + 0.07663781940937042 + ], + "credit_map": [ + 0.003009412670508027, + 0.00449851481243968, + 0.0048040179535746574, + 0.00560237281024456, + 0.00585372606292367, + 0.007201790809631348, + 0.007226070389151573, + 0.005964350420981646, + 0.004448136780411005, + 0.0061514838598668575, + 0.00702779833227396, + 0.006494041066616774, + 0.008419911377131939, + 0.005284998100250959, + 0.007952193729579449, + 0.008123232051730156, + 0.014697184786200523, + 0.007403336465358734, + 0.17639347910881042, + 0.025910677388310432, + 0.00912313349545002, + 0.020538214594125748, + 0.023072456941008568, + 0.1804511994123459, + 0.0033098929561674595, + 0.0027472679503262043, + 0.004197745118290186, + 0.004520016256719828, + 0.004137833137065172, + 0.0038298824802041054, + 0.006439886521548033, + 0.005745831876993179, + 0.003950310871005058, + 0.004241129849106073, + 0.006471958477050066, + 0.0058928076177835464, + 0.005002653691917658, + 0.004697962664067745, + 0.004202218726277351, + 0.004267632961273193, + 0.006310532800853252, + 0.004929948598146439, + 0.005088716745376587, + 0.007239784114062786, + 0.0054854778572916985, + 0.0038274747785180807, + 0.0043893177062273026, + 0.004367159213870764, + 0.00865590013563633, + 0.009541653096675873, + 0.010666616261005402, + 0.011433441191911697, + 0.012550562620162964, + 0.009090820327401161, + 0.01586899720132351, + 0.00999708566814661, + 0.005724385380744934, + 0.008033857680857182, + 0.006923814304172993, + 0.011657727882266045, + 0.009131459519267082, + 0.009327154606580734, + 0.015004469081759453, + 0.009312075562775135, + 0.006325575057417154, + 0.01062529906630516, + 0.012557273730635643, + 0.011229965835809708, + 0.01173823419958353, + 0.008781122043728828, + 0.015221241861581802, + 0.013018229976296425, + 0.012296125292778015, + 0.00687818368896842, + 0.009312796406447887, + 0.009374026209115982, + 0.007635161280632019, + 0.012591663748025894, + 0.008742004632949829, + 0.009807848371565342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.03813860937952995 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669084906578064 + ], + "opposite_logprob_tokens": [ + -3.0486690998077393 + ], + "logprob_reward": -0.048669084906578064, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669084906578064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22910833358764648, + 0.12931698560714722, + 0.13679735362529755, + 0.06821485608816147, + 0.06035732850432396, + 0.05102338641881943, + 0.12461906671524048, + 0.03521370887756348, + 0.0500645786523819, + 0.11528442800045013 + ], + "credit_map": [ + 0.014111531898379326, + 0.02772359363734722, + 0.020878762006759644, + 0.023427249863743782, + 0.031167009845376015, + 0.029447441920638084, + 0.04575180262327194, + 0.03660093620419502, + 0.01073793787509203, + 0.01041122991591692, + 0.01837477646768093, + 0.0164677444845438, + 0.019456274807453156, + 0.016162533313035965, + 0.021923063322901726, + 0.01578342542052269, + 0.008929964154958725, + 0.01028048899024725, + 0.026349002495408058, + 0.017388345673680305, + 0.017108524218201637, + 0.016161274164915085, + 0.013353238813579082, + 0.027226509526371956, + 0.00564994802698493, + 0.007815874181687832, + 0.008726892992854118, + 0.010065310634672642, + 0.006932309828698635, + 0.00704558240249753, + 0.012991692870855331, + 0.0089872470125556, + 0.0044363755732774734, + 0.005468589253723621, + 0.011020012199878693, + 0.012147637084126472, + 0.005827869288623333, + 0.007183589041233063, + 0.006722724996507168, + 0.007550529204308987, + 0.006597572937607765, + 0.005232927389442921, + 0.006440356373786926, + 0.009003122337162495, + 0.004776726942509413, + 0.005231055431067944, + 0.008473701775074005, + 0.005267926957458258, + 0.013877443969249725, + 0.012770969420671463, + 0.014945022761821747, + 0.014129212126135826, + 0.019202426075935364, + 0.013464552350342274, + 0.02126302756369114, + 0.014966421760618687, + 0.003447899827733636, + 0.0033009578473865986, + 0.0031629952136427164, + 0.00598650798201561, + 0.0045833284966647625, + 0.00396786630153656, + 0.006635118741542101, + 0.0041290330700576305, + 0.004585625603795052, + 0.007553871255367994, + 0.008327757008373737, + 0.005830750800669193, + 0.005709752906113863, + 0.00430719181895256, + 0.006131337024271488, + 0.007618293631821871, + 0.014323573559522629, + 0.011292969807982445, + 0.015670739114284515, + 0.017286496236920357, + 0.01107682753354311, + 0.014019692316651344, + 0.018954606726765633, + 0.012659531086683273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.048669084906578064 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639340288937092 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639340288937092, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639340288937092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11472654342651367, + 0.05819837376475334, + 0.17191825807094574, + 0.18007975816726685, + 0.07655426859855652, + 0.04767191782593727, + 0.09365905076265335, + 0.02600957453250885, + 0.17760798335075378, + 0.05357424542307854 + ], + "credit_map": [ + 0.008221675641834736, + 0.009516194462776184, + 0.011866824701428413, + 0.01156060304492712, + 0.01441691443324089, + 0.011294182389974594, + 0.02663976140320301, + 0.021210400387644768, + 0.005625172518193722, + 0.005405622534453869, + 0.007271803915500641, + 0.006001340225338936, + 0.005923254415392876, + 0.006299240048974752, + 0.01279172021895647, + 0.008880224078893661, + 0.006090785842388868, + 0.005689059849828482, + 0.0612071231007576, + 0.011794226244091988, + 0.006127669010311365, + 0.00912052858620882, + 0.00876903161406517, + 0.06311984360218048, + 0.013744735158979893, + 0.018791360780596733, + 0.02324509248137474, + 0.025724880397319794, + 0.02406773902475834, + 0.020243678241968155, + 0.02926456741988659, + 0.024997718632221222, + 0.005878513678908348, + 0.006699298042804003, + 0.015887292101979256, + 0.017410743981599808, + 0.007579599041491747, + 0.006386488676071167, + 0.008277365937829018, + 0.00843497272580862, + 0.005789714399725199, + 0.004369038622826338, + 0.006490449421107769, + 0.00920680072158575, + 0.004182322882115841, + 0.0046457224525511265, + 0.007447363808751106, + 0.005540512036532164, + 0.011839000508189201, + 0.008512025699019432, + 0.009794040583074093, + 0.0106217535212636, + 0.012207464314997196, + 0.008830709382891655, + 0.02053380198776722, + 0.011320260353386402, + 0.0018844762817025185, + 0.001910473802126944, + 0.0017410411965101957, + 0.002810956910252571, + 0.002407385967671871, + 0.0036010798066854477, + 0.007070080377161503, + 0.004584082867950201, + 0.014510714448988438, + 0.029578087851405144, + 0.02476765774190426, + 0.01935383304953575, + 0.026244712993502617, + 0.015974394977092743, + 0.024285949766635895, + 0.022892644628882408, + 0.007329599000513554, + 0.0051856874488294125, + 0.007062028162181377, + 0.007362722419202328, + 0.005394000560045242, + 0.008289267309010029, + 0.006519347429275513, + 0.006431601010262966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.008639340288937092 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638749830424786 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638749830424786, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638749830424786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0708390474319458, + 0.050770051777362823, + 0.3888227343559265, + 0.06281892210245132, + 0.12252243608236313, + 0.02789747342467308, + 0.08215770125389099, + 0.0419364832341671, + 0.05665689706802368, + 0.09557824581861496 + ], + "credit_map": [ + 0.006030124146491289, + 0.008930277079343796, + 0.0066263871267437935, + 0.009144536219537258, + 0.008033119142055511, + 0.010228586383163929, + 0.012491940520703793, + 0.009354072622954845, + 0.00519185746088624, + 0.004332089796662331, + 0.0052264672704041, + 0.005700456909835339, + 0.0073421117849648, + 0.0051248264499008656, + 0.00892729964107275, + 0.00892493687570095, + 0.012991294264793396, + 0.004592197481542826, + 0.15093587338924408, + 0.02218710631132126, + 0.007411783095449209, + 0.01830514706671238, + 0.01979064755141735, + 0.15260867774486542, + 0.00544298579916358, + 0.008617537096142769, + 0.007359533105045557, + 0.008507223799824715, + 0.008992619812488556, + 0.006790638901293278, + 0.00794230680912733, + 0.009166079573333263, + 0.008809328079223633, + 0.011263254098594189, + 0.024085363373160362, + 0.02877720817923546, + 0.010742847807705402, + 0.010556346736848354, + 0.012918884865939617, + 0.015369202941656113, + 0.002988086547702551, + 0.0026431004516780376, + 0.004185724537819624, + 0.0064700269140303135, + 0.0026134836953133345, + 0.002434065565466881, + 0.003459322964772582, + 0.0031036632135510445, + 0.00815451331436634, + 0.009785402566194534, + 0.016440633684396744, + 0.014801686629652977, + 0.009891810826957226, + 0.007566204760223627, + 0.007769865915179253, + 0.007747584953904152, + 0.003986254800111055, + 0.004422127269208431, + 0.0035312685649842024, + 0.005435096565634012, + 0.0067859431728720665, + 0.005802201572805643, + 0.007356254383921623, + 0.004617335740476847, + 0.004001058638095856, + 0.006978352088481188, + 0.00592438131570816, + 0.007754793856292963, + 0.010653141885995865, + 0.005621211137622595, + 0.009998435154557228, + 0.005725522991269827, + 0.012579528614878654, + 0.00960042979568243, + 0.010108865797519684, + 0.0100266607478261, + 0.009312117472290993, + 0.020849991589784622, + 0.010322364047169685, + 0.012778287753462791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.008638749830424786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, looking upward from the ground, the camera tilts steadily upward, closely following the foot of a massive mechanical armor. This foot is composed of battle-worn, dark gray metal armor, its surface covered in scratches and intricate wiring. A massive thruster nozzle at the center of its sole suddenly unleashes a stream of blindingly hot blue-white plasma. This destructive energy instantly pulverizes the concrete below, unleashing a torrent of dense smoke and rubble of varying sizes, which scatters in all directions in a powerful shockwave. Driven by immense force, the behemoth rises vertically in a steady and unstoppable manner. The blazing air creates distinct ripples in the air, casting a dazzling light on the swirling cloud of smoke and dust and the armor's rough metal legs, creating a violent flicker of light and shadow. The entire picture is filled with overwhelming power and apocalyptic grandeur, with tones dominated by cold gray and metallic colors. Only the dazzling blue and white of the energy beam becomes the only color core in the entire scene, full of destructive vitality.", + "seed": 23222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_0/index_6713_seed_23222_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": -0.03816276788711548, + "vlm_logprob": -0.03816276788711548, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03816276788711548, + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11642354726791382 + ], + "opposite_logprob_tokens": [ + -2.2414236068725586 + ], + "logprob_reward": -0.11642354726791382, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11642354726791382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08468171209096909, + 0.06004386022686958, + 0.039952103048563004, + 0.23466990888118744, + 0.05810469761490822, + 0.05122504010796547, + 0.2557130455970764, + 0.05339694023132324, + 0.13458400964736938, + 0.027628641575574875 + ], + "credit_map": [ + 0.007414944004267454, + 0.010971436277031898, + 0.009634529240429401, + 0.01569453440606594, + 0.010968673042953014, + 0.011140813119709492, + 0.01182213518768549, + 0.007034657057374716, + 0.006658692378550768, + 0.007896755822002888, + 0.006976219359785318, + 0.009470229037106037, + 0.007491311989724636, + 0.005751732736825943, + 0.008276800625026226, + 0.007522122003138065, + 0.003057637484744191, + 0.0029561768751591444, + 0.007332415319979191, + 0.006109272595494986, + 0.0037509652320295572, + 0.006876044441014528, + 0.005717949941754341, + 0.004151643719524145, + 0.011972570791840553, + 0.008853674866259098, + 0.06766685098409653, + 0.013430619612336159, + 0.010836075060069561, + 0.004285512492060661, + 0.09254393726587296, + 0.025080682709813118, + 0.005197786260396242, + 0.008406112901866436, + 0.00905703566968441, + 0.007452813908457756, + 0.006333759054541588, + 0.01139578316360712, + 0.00581548735499382, + 0.0044459193013608456, + 0.004434448666870594, + 0.005280298180878162, + 0.008393257856369019, + 0.011996123008430004, + 0.004948817193508148, + 0.0039163511246442795, + 0.005346341524273157, + 0.0069094086065888405, + 0.005537741351872683, + 0.009639776311814785, + 0.046869926154613495, + 0.06055719032883644, + 0.008870722725987434, + 0.005880182608962059, + 0.10596147179603577, + 0.012396060861647129, + 0.0039020583499222994, + 0.007411192171275616, + 0.007725447416305542, + 0.007461232133209705, + 0.009628240950405598, + 0.005473253782838583, + 0.00620888639241457, + 0.005586630664765835, + 0.012560784816741943, + 0.01230697799474001, + 0.026283496990799904, + 0.018246907740831375, + 0.010874679312109947, + 0.022729014977812767, + 0.01724446751177311, + 0.014337687753140926, + 0.005515210796147585, + 0.0031035393476486206, + 0.0032410062849521637, + 0.0026319846510887146, + 0.0026651923544704914, + 0.005106125492602587, + 0.002976437099277973, + 0.0023891485761851072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.11642354726791382 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219275675714016 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219275675714016, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219275675714016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06280264258384705, + 0.09798014163970947, + 0.06549176573753357, + 0.22534461319446564, + 0.02296784147620201, + 0.07586213946342468, + 0.18764187395572662, + 0.056115880608558655, + 0.131740540266037, + 0.07405249774456024 + ], + "credit_map": [ + 0.005013774149119854, + 0.006889022886753082, + 0.006995076779276133, + 0.009604651480913162, + 0.009689626283943653, + 0.006999820005148649, + 0.010522674769163132, + 0.007087997626513243, + 0.010335227474570274, + 0.013021588325500488, + 0.010562923736870289, + 0.013139640912413597, + 0.012239706702530384, + 0.01326787006109953, + 0.012023414485156536, + 0.013389777392148972, + 0.003919636365026236, + 0.006710261572152376, + 0.006671815644949675, + 0.006281261797994375, + 0.014038464985787868, + 0.014779380522668362, + 0.006298973690718412, + 0.006791976746171713, + 0.006900148466229439, + 0.005597018636763096, + 0.06658586859703064, + 0.011234205216169357, + 0.010401651263237, + 0.00330668268725276, + 0.09373890608549118, + 0.02758014015853405, + 0.0031219159718602896, + 0.0021180675830692053, + 0.003931810148060322, + 0.002907062880694866, + 0.0018733066972345114, + 0.0032552988268435, + 0.0030087619088590145, + 0.0027516193222254515, + 0.004552039783447981, + 0.006383935920894146, + 0.02293025143444538, + 0.017536954954266548, + 0.005469246301800013, + 0.004413355607539415, + 0.0070418319664895535, + 0.00753452954813838, + 0.004593220539391041, + 0.006256638560444117, + 0.03178263083100319, + 0.04576494172215462, + 0.007838975638151169, + 0.006086835637688637, + 0.07416661828756332, + 0.01115201972424984, + 0.0037111518904566765, + 0.005578927230089903, + 0.011510788463056087, + 0.007257760502398014, + 0.007373024243861437, + 0.008849204517900944, + 0.005503460764884949, + 0.00633156206458807, + 0.012089828960597515, + 0.011582805775105953, + 0.02562055177986622, + 0.016330385580658913, + 0.011164245195686817, + 0.02046913281083107, + 0.01857498288154602, + 0.015908606350421906, + 0.013362565077841282, + 0.009910405613481998, + 0.008479821495711803, + 0.007374223321676254, + 0.008006097748875618, + 0.012608027085661888, + 0.007962263189256191, + 0.006349103059619665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.014219275675714016 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812438413500786 + ], + "opposite_logprob_tokens": [ + -3.5298123359680176 + ], + "logprob_reward": -0.029812438413500786, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812438413500786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06738323718309402, + 0.1380273848772049, + 0.1607569009065628, + 0.2252701222896576, + 0.03754343092441559, + 0.04363324120640755, + 0.11336065828800201, + 0.09479306638240814, + 0.05097000300884247, + 0.06826190650463104 + ], + "credit_map": [ + 0.0052680387161672115, + 0.006431468762457371, + 0.013757937587797642, + 0.011981392279267311, + 0.006384219042956829, + 0.005792801734060049, + 0.009782074019312859, + 0.007985304109752178, + 0.019667742773890495, + 0.017824534326791763, + 0.014085868373513222, + 0.021423306316137314, + 0.017558490857481956, + 0.013197965919971466, + 0.016391431912779808, + 0.01787804253399372, + 0.005527768284082413, + 0.0094464011490345, + 0.015593959018588066, + 0.015948904678225517, + 0.01052054762840271, + 0.02484644018113613, + 0.050192784518003464, + 0.028680087998509407, + 0.012038199231028557, + 0.009870790876448154, + 0.06817319989204407, + 0.01191676501184702, + 0.011470857076346874, + 0.00576099194586277, + 0.08406582474708557, + 0.02197350189089775, + 0.004830662626773119, + 0.0061307502910494804, + 0.00460011838003993, + 0.004957275930792093, + 0.004643235355615616, + 0.005495250225067139, + 0.003997367341071367, + 0.0028887705411762, + 0.004020381718873978, + 0.004691349808126688, + 0.008382802829146385, + 0.008554322645068169, + 0.005715652368962765, + 0.003818792523816228, + 0.005184730049222708, + 0.003265210660174489, + 0.0025178813375532627, + 0.00511354161426425, + 0.007521165069192648, + 0.03873312100768089, + 0.005225867033004761, + 0.004894647281616926, + 0.04391038045287132, + 0.005444047041237354, + 0.006123355124145746, + 0.010871050879359245, + 0.013364480808377266, + 0.011497209779918194, + 0.012417657300829887, + 0.009627318941056728, + 0.01822921261191368, + 0.012662786990404129, + 0.005182705819606781, + 0.0042157103307545185, + 0.009193696081638336, + 0.006748146377503872, + 0.004355018958449364, + 0.007598587311804295, + 0.007721714209765196, + 0.005954425316303968, + 0.010788322426378727, + 0.009587639942765236, + 0.008307705633342266, + 0.006010890938341618, + 0.008814897388219833, + 0.01023143157362938, + 0.00836144294589758, + 0.006159572396427393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.029812438413500786 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577052608132362 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.020577052608132362, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577052608132362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13101810216903687, + 0.05923391133546829, + 0.05548293888568878, + 0.05559780076146126, + 0.03624558821320534, + 0.21987022459506989, + 0.13395680487155914, + 0.13044261932373047, + 0.08901593834161758, + 0.08913600444793701 + ], + "credit_map": [ + 0.01069349143654108, + 0.01332771684974432, + 0.018967149779200554, + 0.017940299585461617, + 0.016169505193829536, + 0.011889063753187656, + 0.022386129945516586, + 0.01964475028216839, + 0.007366842590272427, + 0.008141790516674519, + 0.005055553279817104, + 0.006874012760818005, + 0.00954014714807272, + 0.005755031947046518, + 0.008053210563957691, + 0.008447324857115746, + 0.0037245387211441994, + 0.006187735591083765, + 0.005369159858673811, + 0.0052950503304600716, + 0.009840600192546844, + 0.005336339585483074, + 0.012129993177950382, + 0.007599524687975645, + 0.0075613586232066154, + 0.005090589169412851, + 0.0078888526186347, + 0.007850747555494308, + 0.00644310237839818, + 0.005285928491503, + 0.00817570649087429, + 0.007301516365259886, + 0.0035862810909748077, + 0.005703540053218603, + 0.004409411922097206, + 0.005045205820351839, + 0.0049363067373633385, + 0.005009581800550222, + 0.004642257932573557, + 0.002913001226261258, + 0.016993507742881775, + 0.02589348703622818, + 0.025767607614398003, + 0.05712497979402542, + 0.028363153338432312, + 0.02067188173532486, + 0.02761155739426613, + 0.017444059252738953, + 0.0042111934162676334, + 0.005064282100647688, + 0.04898393526673317, + 0.009378384798765182, + 0.005091829691082239, + 0.004797622095793486, + 0.04890298470854759, + 0.007526579778641462, + 0.007459988817572594, + 0.013739009387791157, + 0.01712668128311634, + 0.012920776382088661, + 0.019910762086510658, + 0.014166323468089104, + 0.027067990973591805, + 0.018051089718937874, + 0.00799617636948824, + 0.008285443298518658, + 0.016961347311735153, + 0.0102770971134305, + 0.00773396622389555, + 0.014027662575244904, + 0.01277097873389721, + 0.010963269509375095, + 0.01486894953995943, + 0.01165720820426941, + 0.009886547923088074, + 0.009177733212709427, + 0.010768357664346695, + 0.014736324548721313, + 0.010785289108753204, + 0.007255595177412033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.020577052608132362 + }, + { + "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 low-angle looking upward from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while following the foot of a massive mechanical armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the armor foot made of dark gray metal with visible scratches and intricate wiring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a thruster nozzle on the sole unleash a stream of blindingly hot blue-white plasma that pulverizes the concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the behemoth rise vertically in a steady manner while creating air ripples and casting dazzling light on the smoke and armor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978152547031641 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.00978152547031641, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978152547031641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07762205600738525, + 0.06621380895376205, + 0.06750594079494476, + 0.1346222311258316, + 0.02536414936184883, + 0.07044454663991928, + 0.10194388031959534, + 0.28316476941108704, + 0.1351623237133026, + 0.03795633092522621 + ], + "credit_map": [ + 0.005934339016675949, + 0.009742340072989464, + 0.009519689716398716, + 0.009020231664180756, + 0.006532300729304552, + 0.004489066544920206, + 0.02347400039434433, + 0.00891009159386158, + 0.004582935944199562, + 0.007492940407246351, + 0.00842362456023693, + 0.010112063027918339, + 0.007044619414955378, + 0.008144572377204895, + 0.010825173929333687, + 0.009587881155312061, + 0.0029363774228841066, + 0.0030801724642515182, + 0.0061022210866212845, + 0.00637200940400362, + 0.006570138037204742, + 0.007759518921375275, + 0.021071963012218475, + 0.01361354161053896, + 0.008325742557644844, + 0.006723706144839525, + 0.034008268266916275, + 0.007169663906097412, + 0.007183407433331013, + 0.0029540839605033398, + 0.05130463093519211, + 0.016952726989984512, + 0.0022315816022455692, + 0.003242623060941696, + 0.0031814677640795708, + 0.004216962028294802, + 0.0024447154719382524, + 0.0027476376853883266, + 0.0038210521452128887, + 0.003478108672425151, + 0.004119585268199444, + 0.00392382126301527, + 0.022169342264533043, + 0.014559857547283173, + 0.005269187502563, + 0.005031784530729055, + 0.005899304058402777, + 0.00947166420519352, + 0.003037492511793971, + 0.003961910959333181, + 0.014211746864020824, + 0.028188500553369522, + 0.0038287173956632614, + 0.002776901703327894, + 0.040214285254478455, + 0.005724325310438871, + 0.022043997421860695, + 0.037780433893203735, + 0.04820391163229942, + 0.03759239614009857, + 0.028145311400294304, + 0.025350335985422134, + 0.04799788445234299, + 0.036050498485565186, + 0.01134991180151701, + 0.01315284799784422, + 0.026479452848434448, + 0.014781284146010876, + 0.01237471867352724, + 0.021716687828302383, + 0.018775496631860733, + 0.016531934961676598, + 0.005915127694606781, + 0.005300847347825766, + 0.004293449688702822, + 0.0037463137414306402, + 0.004766554106026888, + 0.006456954870373011, + 0.004285032860934734, + 0.0031920478213578463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.00978152547031641 + } + ], + "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/0000167/group_1/index_tensor([4132, 4132])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_tensor([4132, 4132])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48c4f0174684045272095f234e0ce830d8f568ab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_tensor([4132, 4132])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot sweeps across the highly polished wooden lane surface, playing with a gleaming bowling ball. The ball, a vibrant candy-apple red sphere with a perfect mirror-like finish, spins rapidly on its axis, its rapid motion creating motion blur and streaking distorted reflections of the bright overhead fluorescent lights across its surface. It rushes down the varnished maple planks, which themselves shimmer with a high-gloss finish, casting a secondary, elongated reflection of a red blur. The dark, deep gutters on either side create a stark contrast, focusing the viewer's attention on the pristine white pins arranged in a perfect triangle in the distance. The ball moves with swift and decisive force, gliding frictionlessly along its path in a powerful, straight trajectory, creating a tense, focused atmosphere of pure kinetic energy and imminent impact under the cool, even fluorescent light.", + "seed": 6988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6988_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": -0.1499544091988355, + "vlm_logprob": -0.1499544091988355, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.1499544091988355, + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08149453997612 + ], + "opposite_logprob_tokens": [ + -3.4564945697784424 + ], + "logprob_reward": -0.08149453997612, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.08149453997612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10368592292070389, + 0.0571836419403553, + 0.07432413846254349, + 0.09039675444364548, + 0.10671672970056534, + 0.12612275779247284, + 0.13631854951381683, + 0.058538325130939484, + 0.08084311336278915, + 0.16587010025978088 + ], + "credit_map": [ + 0.005473491735756397, + 0.008210229687392712, + 0.014827882871031761, + 0.012200977653265, + 0.028257906436920166, + 0.005631727632135153, + 0.01609450951218605, + 0.012989199720323086, + 0.004535323474556208, + 0.005488913040608168, + 0.0086152832955122, + 0.013288800604641438, + 0.0055631911382079124, + 0.005534140393137932, + 0.006953539792448282, + 0.007204451132565737, + 0.0069793774746358395, + 0.003812818555161357, + 0.005554897245019674, + 0.005230384413152933, + 0.0060285660438239574, + 0.011658241972327232, + 0.01763550564646721, + 0.017424356192350388, + 0.005973663181066513, + 0.004087044857442379, + 0.007044649217277765, + 0.007745872717350721, + 0.006550998892635107, + 0.004545356146991253, + 0.027349749580025673, + 0.02709941938519478, + 0.0068607269786298275, + 0.011504516936838627, + 0.0161360502243042, + 0.012940130196511745, + 0.013681343756616116, + 0.011413732543587685, + 0.016426194459199905, + 0.017754025757312775, + 0.006431281566619873, + 0.009052061475813389, + 0.027602095156908035, + 0.018951937556266785, + 0.013618945144116879, + 0.01606013812124729, + 0.01776299439370632, + 0.016643311828374863, + 0.0070862616412341595, + 0.0066033522598445415, + 0.01167785469442606, + 0.0438542515039444, + 0.00781870074570179, + 0.013065113686025143, + 0.027144568040966988, + 0.019068431109189987, + 0.0027026685420423746, + 0.00260341539978981, + 0.0051121762953698635, + 0.01862839050590992, + 0.00342870457097888, + 0.0033183752093464136, + 0.012944594956934452, + 0.009800000116229057, + 0.004085710737854242, + 0.007873411290347576, + 0.019713005051016808, + 0.008997119031846523, + 0.011557822115719318, + 0.0051933107897639275, + 0.008759400807321072, + 0.01466333121061325, + 0.017653824761509895, + 0.009478366933763027, + 0.06254785507917404, + 0.02576102875173092, + 0.011336448602378368, + 0.011181153357028961, + 0.015227777883410454, + 0.012683635577559471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.08149453997612 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520796060562134 + ], + "opposite_logprob_tokens": [ + -1.5020796060562134 + ], + "logprob_reward": -0.2520796060562134, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520796060562134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10842207819223404, + 0.025573108345270157, + 0.05245969817042351, + 0.11791759729385376, + 0.06082157790660858, + 0.3209095895290375, + 0.19171151518821716, + 0.02695711888372898, + 0.04193265363574028, + 0.053294990211725235 + ], + "credit_map": [ + 0.006226848345249891, + 0.00966278463602066, + 0.01737329363822937, + 0.02225368469953537, + 0.014185139909386635, + 0.008622962981462479, + 0.01379858423024416, + 0.016298789530992508, + 0.002008866984397173, + 0.002036487450823188, + 0.00462784618139267, + 0.004947867710143328, + 0.001766660250723362, + 0.0036132449749857187, + 0.002568083116784692, + 0.004004056099802256, + 0.0030917213298380375, + 0.002284157555550337, + 0.0032176619861274958, + 0.004787344951182604, + 0.003547512926161289, + 0.004459309857338667, + 0.015734411776065826, + 0.015337586402893066, + 0.006717019248753786, + 0.006363787688314915, + 0.011330456472933292, + 0.007484475150704384, + 0.00925765186548233, + 0.0063044424168765545, + 0.03497893735766411, + 0.035480841994285583, + 0.004015241749584675, + 0.0074541098438203335, + 0.00796015840023756, + 0.00860055722296238, + 0.009021660313010216, + 0.004437228199094534, + 0.009769822470843792, + 0.009562814608216286, + 0.011633540503680706, + 0.01069179829210043, + 0.06755095720291138, + 0.11429761350154877, + 0.01345905289053917, + 0.018511056900024414, + 0.06198575720191002, + 0.022779861465096474, + 0.008386814035475254, + 0.00970001146197319, + 0.012803331948816776, + 0.05629115179181099, + 0.01173311471939087, + 0.021904844790697098, + 0.032994307577610016, + 0.03789797052741051, + 0.001961258240044117, + 0.0016423965571448207, + 0.0023835699539631605, + 0.003827480599284172, + 0.001975563820451498, + 0.0027151950635015965, + 0.004649937152862549, + 0.007801723200827837, + 0.002035548910498619, + 0.0034770453348755836, + 0.011806433089077473, + 0.0046644918620586395, + 0.00445826118811965, + 0.002629978582262993, + 0.004114833660423756, + 0.008746068924665451, + 0.003943230956792831, + 0.007362480740994215, + 0.007072424050420523, + 0.005634678062051535, + 0.009915758855640888, + 0.005524979904294014, + 0.005978806409984827, + 0.007862645201385021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.2520796060562134 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249610492028296 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.00249610492028296, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249610492028296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059420451521873474, + 0.03330153971910477, + 0.021044012159109116, + 0.13322190940380096, + 0.1254483014345169, + 0.2008579969406128, + 0.1203935444355011, + 0.11470536887645721, + 0.10856597870588303, + 0.08304090797901154 + ], + "credit_map": [ + 0.002755895024165511, + 0.004171844571828842, + 0.006725890561938286, + 0.007889033295214176, + 0.015999125316739082, + 0.004140167497098446, + 0.009533333592116833, + 0.008205164223909378, + 0.00198549241758883, + 0.0021664509549736977, + 0.004430308006703854, + 0.00586312310770154, + 0.0033928407356142998, + 0.0035510407760739326, + 0.004290828946977854, + 0.007621454540640116, + 0.0015266534173861146, + 0.001571420463733375, + 0.002991717541590333, + 0.003108583390712738, + 0.002333298558369279, + 0.0023493727203458548, + 0.0036856273654848337, + 0.0034773380029946566, + 0.006641837302595377, + 0.006216851528733969, + 0.010960996150970459, + 0.013560470193624496, + 0.009099306538701057, + 0.006467333063483238, + 0.038918059319257736, + 0.041357044130563736, + 0.007963902316987514, + 0.010301832109689713, + 0.017003100365400314, + 0.013790356926620007, + 0.01421386655420065, + 0.011417565867304802, + 0.018746396526694298, + 0.03201127424836159, + 0.007677969988435507, + 0.007133597042411566, + 0.046538352966308594, + 0.07183663547039032, + 0.007583152502775192, + 0.01051450427621603, + 0.034663572907447815, + 0.01491022389382124, + 0.005467053968459368, + 0.006149204447865486, + 0.010917693376541138, + 0.03260841220617294, + 0.007448336575180292, + 0.017209801822900772, + 0.02085188776254654, + 0.019741155207157135, + 0.004761972464621067, + 0.009196028113365173, + 0.008892850950360298, + 0.028394026681780815, + 0.010935872793197632, + 0.008955094031989574, + 0.02080250158905983, + 0.022767020389437675, + 0.006116489879786968, + 0.009690165519714355, + 0.024943338707089424, + 0.014268264174461365, + 0.00840866006910801, + 0.008812624961137772, + 0.013708177022635937, + 0.022618262097239494, + 0.008685545064508915, + 0.006807758938521147, + 0.02012484334409237, + 0.014208548702299595, + 0.010565448552370071, + 0.005804087035357952, + 0.007715936284512281, + 0.009128741919994354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.00249610492028296 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030611604452133 + ], + "opposite_logprob_tokens": [ + -2.3503060340881348 + ], + "logprob_reward": -0.10030611604452133, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030611604452133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07314793765544891, + 0.018188392743468285, + 0.030114633962512016, + 0.25974521040916443, + 0.05294116958975792, + 0.21082796156406403, + 0.17741337418556213, + 0.04114312306046486, + 0.06119228154420853, + 0.07528593391180038 + ], + "credit_map": [ + 0.0044760205782949924, + 0.006796690635383129, + 0.009108923375606537, + 0.01103253848850727, + 0.014370670542120934, + 0.003762913169339299, + 0.009869327768683434, + 0.013730846345424652, + 0.0015035602264106274, + 0.0012939282460138202, + 0.0021902944426983595, + 0.0029937750659883022, + 0.0032086900901049376, + 0.0014409349532797933, + 0.002836675150319934, + 0.0027205334044992924, + 0.002780064009130001, + 0.0015277218772098422, + 0.0031831227242946625, + 0.0032980586402118206, + 0.0026480238884687424, + 0.003867289749905467, + 0.006726868450641632, + 0.006083482876420021, + 0.013186546973884106, + 0.01118108443915844, + 0.021072333678603172, + 0.02051454968750477, + 0.017280535772442818, + 0.013836819678544998, + 0.08056461066007614, + 0.08210869878530502, + 0.002981406869366765, + 0.006028043106198311, + 0.007926085032522678, + 0.005142305511981249, + 0.006958880927413702, + 0.004683324601501226, + 0.007802519481629133, + 0.011418596841394901, + 0.008079025894403458, + 0.010048329830169678, + 0.031043224036693573, + 0.07698877155780792, + 0.012254943139851093, + 0.012038967572152615, + 0.035411231219768524, + 0.024963460862636566, + 0.009169433265924454, + 0.009007120504975319, + 0.016850851476192474, + 0.043950341641902924, + 0.011713157407939434, + 0.023227214813232422, + 0.02672254852950573, + 0.03677268326282501, + 0.00268072122707963, + 0.0036011207848787308, + 0.003925020806491375, + 0.0065834722481667995, + 0.005363744217902422, + 0.005083086900413036, + 0.006642975844442844, + 0.007262978237122297, + 0.0029139358084648848, + 0.004108443856239319, + 0.01773015223443508, + 0.005006743129342794, + 0.00396212562918663, + 0.004684306215494871, + 0.005776124075055122, + 0.017010442912578583, + 0.0058739748783409595, + 0.0074936021119356155, + 0.016188353300094604, + 0.006933019962161779, + 0.011743240058422089, + 0.0071297502145171165, + 0.00919171329587698, + 0.010732275433838367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.10030611604452133 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3133956789970398 + ], + "opposite_logprob_tokens": [ + -1.3133957386016846 + ], + "logprob_reward": -0.3133956789970398, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3133956789970398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08194080740213394, + 0.017463019117712975, + 0.03976175934076309, + 0.09345689415931702, + 0.10312402993440628, + 0.10618683695793152, + 0.047281552106142044, + 0.06028208136558533, + 0.15545639395713806, + 0.2950466275215149 + ], + "credit_map": [ + 0.005530714523047209, + 0.006735669914633036, + 0.011738949455320835, + 0.016461294144392014, + 0.008118978701531887, + 0.0070655555464327335, + 0.010282880626618862, + 0.01600676216185093, + 0.0012409994378685951, + 0.0012420538114383817, + 0.0027824249118566513, + 0.0030718680936843157, + 0.0017984391888603568, + 0.0017304993234574795, + 0.002691867295652628, + 0.002904867986217141, + 0.004208771977573633, + 0.0021038069389760494, + 0.0032520126551389694, + 0.004456997383385897, + 0.0034118525218218565, + 0.005017112474888563, + 0.008567465469241142, + 0.008743741549551487, + 0.004378645680844784, + 0.003333142027258873, + 0.008738799020648003, + 0.009705214761197567, + 0.005436906591057777, + 0.004005841910839081, + 0.027777330949902534, + 0.030081015080213547, + 0.008486082777380943, + 0.00727522186934948, + 0.014398724772036076, + 0.010807630605995655, + 0.01014265138655901, + 0.007723089773207903, + 0.015683673322200775, + 0.028606947511434555, + 0.005708925426006317, + 0.00506662018597126, + 0.03123265504837036, + 0.018869221210479736, + 0.0064517720602452755, + 0.010103688575327396, + 0.014846614561975002, + 0.01390734314918518, + 0.0028921456541866064, + 0.0027247723191976547, + 0.005215165205299854, + 0.010713810101151466, + 0.0036427790764719248, + 0.005471062380820513, + 0.010532557033002377, + 0.006089258007705212, + 0.0040008388459682465, + 0.0040174308232963085, + 0.007530920207500458, + 0.011992359533905983, + 0.0038467401172965765, + 0.005614735651761293, + 0.011323118582367897, + 0.011955934576690197, + 0.005926912650465965, + 0.0079817995429039, + 0.0449305884540081, + 0.011110902763903141, + 0.015273714438080788, + 0.012785712257027626, + 0.012620742432773113, + 0.04482602700591087, + 0.03460995480418205, + 0.022990871220827103, + 0.06707514077425003, + 0.04763270169496536, + 0.03525378182530403, + 0.025323960930109024, + 0.02934988960623741, + 0.03281034901738167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.3133956789970398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot sweeps across the highly polished wooden lane surface, playing with a gleaming bowling ball. The ball, a vibrant candy-apple red sphere with a perfect mirror-like finish, spins rapidly on its axis, its rapid motion creating motion blur and streaking distorted reflections of the bright overhead fluorescent lights across its surface. It rushes down the varnished maple planks, which themselves shimmer with a high-gloss finish, casting a secondary, elongated reflection of a red blur. The dark, deep gutters on either side create a stark contrast, focusing the viewer's attention on the pristine white pins arranged in a perfect triangle in the distance. The ball moves with swift and decisive force, gliding frictionlessly along its path in a powerful, straight trajectory, creating a tense, focused atmosphere of pure kinetic energy and imminent impact under the cool, even fluorescent light.", + "seed": 6989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6989_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": -0.2143131542019546, + "vlm_logprob": -0.2143131542019546, + "vlm_yes_no_margin": 2.175000011920929, + "vlm_token_credit": -0.2143131542019546, + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18487071990966797 + ], + "opposite_logprob_tokens": [ + -2.184870719909668 + ], + "logprob_reward": -0.18487071990966797, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.18487071990966797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11094824224710464, + 0.1197984516620636, + 0.05144040286540985, + 0.4251633286476135, + 0.02892068587243557, + 0.030034229159355164, + 0.03076748177409172, + 0.032291337847709656, + 0.11698713153600693, + 0.05364874005317688 + ], + "credit_map": [ + 0.008341909386217594, + 0.012155679054558277, + 0.03070738911628723, + 0.02405138686299324, + 0.006994370836764574, + 0.0070359609089791775, + 0.014743080362677574, + 0.006918465252965689, + 0.015130366198718548, + 0.016431599855422974, + 0.024185622110962868, + 0.02248484082520008, + 0.009711087681353092, + 0.006807377561926842, + 0.013714462518692017, + 0.011333095841109753, + 0.004903233144432306, + 0.01103217527270317, + 0.006936702877283096, + 0.00722788879647851, + 0.0061571947298944, + 0.002820888999849558, + 0.008119181729853153, + 0.004243140108883381, + 0.015676958486437798, + 0.014066276140511036, + 0.1833101511001587, + 0.17748582363128662, + 0.009343137964606285, + 0.004310628864914179, + 0.01382837351411581, + 0.007141999900341034, + 0.0031694553326815367, + 0.0036835544742643833, + 0.004973667208105326, + 0.004134785383939743, + 0.0032975056674331427, + 0.003056707326322794, + 0.0039857011288404465, + 0.0026193077210336924, + 0.004101903643459082, + 0.005474272649735212, + 0.006284593604505062, + 0.004447392653673887, + 0.0040442016907036304, + 0.0015472319209948182, + 0.002480443799868226, + 0.0016541891964152455, + 0.004654062911868095, + 0.005998414009809494, + 0.005705596413463354, + 0.0036052989307790995, + 0.002393182599917054, + 0.0019805303309112787, + 0.0032529584132134914, + 0.0031774381641298532, + 0.00692030880600214, + 0.006119693163782358, + 0.004673979245126247, + 0.003731341101229191, + 0.003505098167806864, + 0.003430094802752137, + 0.002229842822998762, + 0.00168098125141114, + 0.019098637625575066, + 0.014001152478158474, + 0.014913942664861679, + 0.020207080990076065, + 0.014209771528840065, + 0.011047495529055595, + 0.014547890983521938, + 0.008961165323853493, + 0.009248098358511925, + 0.007021957077085972, + 0.007449943106621504, + 0.010535180568695068, + 0.0038832034915685654, + 0.0038329341914504766, + 0.007272352930158377, + 0.004405070561915636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.18487071990966797 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157423615455627 + ], + "opposite_logprob_tokens": [ + -1.7015742063522339 + ], + "logprob_reward": -0.20157423615455627, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157423615455627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13064122200012207, + 0.05700162425637245, + 0.04609956964850426, + 0.344190388917923, + 0.06793221086263657, + 0.0507822148501873, + 0.08724228292703629, + 0.05246112495660782, + 0.05578599497675896, + 0.1078634113073349 + ], + "credit_map": [ + 0.013679806143045425, + 0.01541154645383358, + 0.02947499044239521, + 0.02366683818399906, + 0.012556001543998718, + 0.008463455364108086, + 0.017302801832556725, + 0.010085788555443287, + 0.007167454808950424, + 0.007836475037038326, + 0.012109475210309029, + 0.011509240604937077, + 0.004514261148869991, + 0.00336225563660264, + 0.0053784893825650215, + 0.005123969167470932, + 0.004256489221006632, + 0.004484964068979025, + 0.01064000278711319, + 0.008003024384379387, + 0.004262933041900396, + 0.002037026919424534, + 0.008371299132704735, + 0.004043830092996359, + 0.015088138170540333, + 0.01273471862077713, + 0.14270395040512085, + 0.14177247881889343, + 0.009227885864675045, + 0.004320647567510605, + 0.01153324544429779, + 0.006809295155107975, + 0.008441850543022156, + 0.007351275999099016, + 0.015180517919361591, + 0.008548560552299023, + 0.010879923589527607, + 0.0054163034074008465, + 0.007752299774438143, + 0.00436148326843977, + 0.006165382917970419, + 0.0066004362888634205, + 0.01188971009105444, + 0.006875814404338598, + 0.005852727685123682, + 0.003987746778875589, + 0.00513059925287962, + 0.004279796499758959, + 0.015504228882491589, + 0.01579073816537857, + 0.0167696513235569, + 0.010051459074020386, + 0.008271027356386185, + 0.005769122391939163, + 0.008231844753026962, + 0.0068542188964784145, + 0.010132886469364166, + 0.00766276940703392, + 0.008881660178303719, + 0.005347817204892635, + 0.007068549748510122, + 0.004264817573130131, + 0.004564649425446987, + 0.00453797634691, + 0.009440368972718716, + 0.006345417816191912, + 0.005358307156711817, + 0.008821563795208931, + 0.005550259258598089, + 0.007382573094218969, + 0.007769827730953693, + 0.005117679014801979, + 0.013283856213092804, + 0.022666871547698975, + 0.010834408923983574, + 0.01883012428879738, + 0.008269048295915127, + 0.012508629821240902, + 0.01244462002068758, + 0.009025845676660538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.20157423615455627 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410571089014411 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.00410571089014411, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410571089014411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15105709433555603, + 0.20695246756076813, + 0.09652552753686905, + 0.13012486696243286, + 0.05638621002435684, + 0.055526796728372574, + 0.10924703627824783, + 0.04224344715476036, + 0.049271777272224426, + 0.10266483575105667 + ], + "credit_map": [ + 0.012317444197833538, + 0.01637323386967182, + 0.0321832001209259, + 0.0344102680683136, + 0.009635498747229576, + 0.008136768825352192, + 0.02853785827755928, + 0.00946281012147665, + 0.023161577060818672, + 0.04087641462683678, + 0.026718175038695335, + 0.02679922617971897, + 0.029142912477254868, + 0.020870987325906754, + 0.02393076941370964, + 0.01545239333063364, + 0.009712738916277885, + 0.020522089675068855, + 0.0146839190274477, + 0.013343919068574905, + 0.012697218917310238, + 0.005428323522210121, + 0.012389927171170712, + 0.007747378200292587, + 0.010930093936622143, + 0.015325085259974003, + 0.026700714603066444, + 0.037966448813676834, + 0.010534425266087055, + 0.007999944500625134, + 0.011850251816213131, + 0.008817890658974648, + 0.005059219431132078, + 0.005693551618605852, + 0.008930742740631104, + 0.006989858578890562, + 0.010499781928956509, + 0.006671308074146509, + 0.007188171148300171, + 0.0053535704500973225, + 0.006731436122208834, + 0.009133081883192062, + 0.012013682164251804, + 0.007713695522397757, + 0.006266366690397263, + 0.0053578694351017475, + 0.004399081226438284, + 0.003911580890417099, + 0.015354030765593052, + 0.022096604108810425, + 0.01927606761455536, + 0.011661738157272339, + 0.014075462706387043, + 0.007494947873055935, + 0.008851162157952785, + 0.010437019169330597, + 0.009088332764804363, + 0.008038881234824657, + 0.004168310668319464, + 0.0045442841947078705, + 0.005998473148792982, + 0.003602945478633046, + 0.0037115421146154404, + 0.003090675687417388, + 0.008900373242795467, + 0.006085806060582399, + 0.005838611628860235, + 0.005589375272393227, + 0.007260703016072512, + 0.004831209313124418, + 0.006067323498427868, + 0.0046983701176941395, + 0.011236021295189857, + 0.015632452443242073, + 0.011515161953866482, + 0.014923580922186375, + 0.020353995263576508, + 0.005812089424580336, + 0.013905322179198265, + 0.00928620994091034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.00410571089014411 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252131849527359 + ], + "opposite_logprob_tokens": [ + -1.5021318197250366 + ], + "logprob_reward": -0.252131849527359, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252131849527359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14693795144557953, + 0.07905929535627365, + 0.09780960530042648, + 0.28141388297080994, + 0.04157205671072006, + 0.05411263927817345, + 0.04872642830014229, + 0.06709359586238861, + 0.0361202135682106, + 0.14715437591075897 + ], + "credit_map": [ + 0.013736763969063759, + 0.021440181881189346, + 0.02431698702275753, + 0.018085312098264694, + 0.013875375501811504, + 0.01835356093943119, + 0.02421223744750023, + 0.012917522341012955, + 0.00952232163399458, + 0.012771246954798698, + 0.011611178517341614, + 0.011176668107509613, + 0.013164469040930271, + 0.0060069262981414795, + 0.009256681427359581, + 0.005549803376197815, + 0.011134780943393707, + 0.017139334231615067, + 0.020291369408369064, + 0.013767430558800697, + 0.009209652431309223, + 0.003733098041266203, + 0.013187672011554241, + 0.009346263483166695, + 0.01068140473216772, + 0.010457850992679596, + 0.11558858305215836, + 0.1130526140332222, + 0.010169101879000664, + 0.004366135690361261, + 0.01150963082909584, + 0.0055885352194309235, + 0.005588242318481207, + 0.0051627131178975105, + 0.009748215787112713, + 0.005325124599039555, + 0.00503982650116086, + 0.0034881283063441515, + 0.004415646195411682, + 0.0028041608165949583, + 0.0069520059041678905, + 0.00974056776612997, + 0.01476514246314764, + 0.007188254036009312, + 0.004853604827076197, + 0.0034851012751460075, + 0.0035864035598933697, + 0.0035415582824498415, + 0.007071105297654867, + 0.007433006074279547, + 0.009447255171835423, + 0.005059071816504002, + 0.007045522797852755, + 0.004880765452980995, + 0.004220891278237104, + 0.0035688092466443777, + 0.011491721495985985, + 0.009980620816349983, + 0.01194743812084198, + 0.007129897363483906, + 0.009440507739782333, + 0.007142491172999144, + 0.005732403136789799, + 0.0042285132221877575, + 0.0041421460919082165, + 0.0046305106952786446, + 0.004957617726176977, + 0.00588365038856864, + 0.006134754978120327, + 0.0025605279952287674, + 0.004658900201320648, + 0.0031521059572696686, + 0.012337814085185528, + 0.016319403424859047, + 0.014689281582832336, + 0.02444305270910263, + 0.027354927733540535, + 0.009780025109648705, + 0.026570897549390793, + 0.01565898209810257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.252131849527359 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888325452804565 + ], + "opposite_logprob_tokens": [ + -1.0538833141326904 + ], + "logprob_reward": -0.42888325452804565, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42888325452804565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1864473670721054, + 0.08915241807699203, + 0.054567139595746994, + 0.07485803216695786, + 0.04031777381896973, + 0.03340088948607445, + 0.02587958425283432, + 0.08459120243787766, + 0.0721118301153183, + 0.33867380023002625 + ], + "credit_map": [ + 0.01934126578271389, + 0.020563624799251556, + 0.0377015620470047, + 0.029516106471419334, + 0.015507536008954048, + 0.012562917545437813, + 0.033006176352500916, + 0.018248174339532852, + 0.008886918425559998, + 0.013264985755085945, + 0.01656377501785755, + 0.019554980099201202, + 0.00813529547303915, + 0.00679072504863143, + 0.008967770263552666, + 0.0069879707880318165, + 0.004806201905012131, + 0.00603154581040144, + 0.013824081048369408, + 0.008726346306502819, + 0.00649290531873703, + 0.002593038836494088, + 0.00792829878628254, + 0.004164721816778183, + 0.003974223975092173, + 0.005067975725978613, + 0.025595327839255333, + 0.028420528396964073, + 0.0038526190910488367, + 0.0024767497088760138, + 0.003506699576973915, + 0.0019639115780591965, + 0.004240822978317738, + 0.005559708923101425, + 0.010385885834693909, + 0.005246452055871487, + 0.003344951895996928, + 0.0027950957883149385, + 0.005880980286747217, + 0.002863873727619648, + 0.005294834729284048, + 0.006566863972693682, + 0.005767522845417261, + 0.004003151319921017, + 0.004399710334837437, + 0.002195921028032899, + 0.0028076530434191227, + 0.0023652331437915564, + 0.003988809417933226, + 0.0033446054439991713, + 0.004895708989351988, + 0.00351607333868742, + 0.002936940174549818, + 0.001960311084985733, + 0.0027826183941215277, + 0.0024545167107135057, + 0.017098573967814445, + 0.013802249915897846, + 0.012748478911817074, + 0.008074106648564339, + 0.009473572485148907, + 0.009614931419491768, + 0.007606893312186003, + 0.006172392517328262, + 0.01190218050032854, + 0.010585234500467777, + 0.0073081254959106445, + 0.009477886371314526, + 0.008863281458616257, + 0.00872605200856924, + 0.009296390227973461, + 0.00595267815515399, + 0.05604368820786476, + 0.06983120739459991, + 0.025775369256734848, + 0.07577845454216003, + 0.024399155750870705, + 0.030490156263113022, + 0.02484535239636898, + 0.03151041641831398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.42888325452804565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot sweeps across the highly polished wooden lane surface, playing with a gleaming bowling ball. The ball, a vibrant candy-apple red sphere with a perfect mirror-like finish, spins rapidly on its axis, its rapid motion creating motion blur and streaking distorted reflections of the bright overhead fluorescent lights across its surface. It rushes down the varnished maple planks, which themselves shimmer with a high-gloss finish, casting a secondary, elongated reflection of a red blur. The dark, deep gutters on either side create a stark contrast, focusing the viewer's attention on the pristine white pins arranged in a perfect triangle in the distance. The ball moves with swift and decisive force, gliding frictionlessly along its path in a powerful, straight trajectory, creating a tense, focused atmosphere of pure kinetic energy and imminent impact under the cool, even fluorescent light.", + "seed": 6990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6990_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": -0.18972569885663687, + "vlm_logprob": -0.18972569885663687, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.18972569885663687, + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06737210601568222 + ], + "opposite_logprob_tokens": [ + -3.5673720836639404 + ], + "logprob_reward": -0.06737210601568222, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06737210601568222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17055794596672058, + 0.044558070600032806, + 0.03826485574245453, + 0.05996576324105263, + 0.04216863587498665, + 0.0446459986269474, + 0.047853533178567886, + 0.0866587907075882, + 0.33489495515823364, + 0.13043147325515747 + ], + "credit_map": [ + 0.012187985703349113, + 0.027468105778098106, + 0.027938734740018845, + 0.02781584858894348, + 0.018057886511087418, + 0.016355542466044426, + 0.01945035718381405, + 0.021283484995365143, + 0.004165384452790022, + 0.00450693117454648, + 0.0050190528854727745, + 0.004329311661422253, + 0.0053766402415931225, + 0.005684615578502417, + 0.007704296614974737, + 0.007771841250360012, + 0.0020881034433841705, + 0.00894767977297306, + 0.008992936462163925, + 0.007754127960652113, + 0.0032484910916537046, + 0.0017651197267696261, + 0.002746966667473316, + 0.0027214311994612217, + 0.0047960239462554455, + 0.012722558341920376, + 0.008296185173094273, + 0.009215131402015686, + 0.008247783407568932, + 0.00554802967235446, + 0.00641202786937356, + 0.004728024359792471, + 0.0027362543623894453, + 0.005620510317385197, + 0.007371657527983189, + 0.006474480032920837, + 0.0051481230184435844, + 0.002218363108113408, + 0.007223949767649174, + 0.0053752996027469635, + 0.004134334158152342, + 0.007518146652728319, + 0.007869595661759377, + 0.006316345650702715, + 0.0043554967269301414, + 0.0030490148346871138, + 0.005064716096967459, + 0.006338351406157017, + 0.006010974757373333, + 0.005878332536667585, + 0.00667885085567832, + 0.00499738147482276, + 0.0042122104205191135, + 0.004727252293378115, + 0.005521056242287159, + 0.009827475063502789, + 0.006062789354473352, + 0.008606307208538055, + 0.0212128683924675, + 0.01045282743871212, + 0.005805206950753927, + 0.005579083692282438, + 0.007104808464646339, + 0.021834904327988625, + 0.02116396464407444, + 0.026365451514720917, + 0.049474988132715225, + 0.03038184903562069, + 0.08244580775499344, + 0.030557790771126747, + 0.05728352442383766, + 0.03722161054611206, + 0.012818817980587482, + 0.008810125291347504, + 0.024214722216129303, + 0.011808983981609344, + 0.011878818273544312, + 0.009329552762210369, + 0.031095482409000397, + 0.020474979653954506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.06737210601568222 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704976439476013 + ], + "opposite_logprob_tokens": [ + -1.1370497941970825 + ], + "logprob_reward": -0.38704976439476013, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38704976439476013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11876605451107025, + 0.06770732253789902, + 0.04096280783414841, + 0.3509926497936249, + 0.025464264675974846, + 0.023094888776540756, + 0.060955289751291275, + 0.13742271065711975, + 0.052148133516311646, + 0.12248580157756805 + ], + "credit_map": [ + 0.008969897404313087, + 0.018845632672309875, + 0.019675469025969505, + 0.01831970550119877, + 0.011185567826032639, + 0.012379425577819347, + 0.014754765667021275, + 0.014635611325502396, + 0.004846592433750629, + 0.009624636732041836, + 0.008590647019445896, + 0.0072683668695390224, + 0.005758141633123159, + 0.00817100703716278, + 0.013106043450534344, + 0.010341898538172245, + 0.0029652367811650038, + 0.0073157805018126965, + 0.006632813718169928, + 0.006170636974275112, + 0.004533219616860151, + 0.0031963111832737923, + 0.006187275517731905, + 0.003961541224271059, + 0.014654591679573059, + 0.025129182264208794, + 0.13250313699245453, + 0.12457409501075745, + 0.014801755547523499, + 0.011965770274400711, + 0.017085295170545578, + 0.010278874076902866, + 0.001610224018804729, + 0.001934277592226863, + 0.005418031010776758, + 0.003712607314810157, + 0.002662177663296461, + 0.0014926004223525524, + 0.0037282013799995184, + 0.004906149115413427, + 0.0011472307378426194, + 0.0020091196056455374, + 0.0025941592175513506, + 0.005572849418967962, + 0.0019235546933487058, + 0.0014509805478155613, + 0.002284025540575385, + 0.00611297320574522, + 0.009455554187297821, + 0.008707228116691113, + 0.008231112733483315, + 0.008116103708744049, + 0.005424091592431068, + 0.0048514255322515965, + 0.005631143692880869, + 0.010538638569414616, + 0.009137708693742752, + 0.009093211032450199, + 0.047640737146139145, + 0.00911722332239151, + 0.005233081057667732, + 0.006993337534368038, + 0.006865732837468386, + 0.04334169626235962, + 0.004174611996859312, + 0.004385137464851141, + 0.009172937832772732, + 0.005160909611731768, + 0.009282584302127361, + 0.006770607084035873, + 0.0073077492415905, + 0.005893604829907417, + 0.012004563584923744, + 0.010590157471597195, + 0.016212858259677887, + 0.011845093220472336, + 0.014677727594971657, + 0.011307813227176666, + 0.019963065162301064, + 0.02588455006480217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.38704976439476013 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009277330245822668 + ], + "opposite_logprob_tokens": [ + -7.000927925109863 + ], + "logprob_reward": -0.0009277330245822668, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009277330245822668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10391730070114136, + 0.03376537561416626, + 0.01852297969162464, + 0.25870224833488464, + 0.04269281029701233, + 0.01589353382587433, + 0.04572214558720589, + 0.13286980986595154, + 0.20287944376468658, + 0.14503441751003265 + ], + "credit_map": [ + 0.009273404255509377, + 0.014645555056631565, + 0.017180386930704117, + 0.014057597145438194, + 0.010050595737993717, + 0.011119079776108265, + 0.01192204374819994, + 0.015668639913201332, + 0.00241254735738039, + 0.0043692574836313725, + 0.004817663691937923, + 0.0034660256933420897, + 0.004616860765963793, + 0.0038109158631414175, + 0.005606167949736118, + 0.00466593774035573, + 0.0015090079978108406, + 0.003192056203261018, + 0.003639762755483389, + 0.0029183090664446354, + 0.002337375422939658, + 0.0012789550237357616, + 0.0020064262207597494, + 0.0016410864191129804, + 0.010700379498302937, + 0.015725277364253998, + 0.10267042368650436, + 0.09688281267881393, + 0.007192452438175678, + 0.007095179054886103, + 0.010956688784062862, + 0.007479019928723574, + 0.0036626968067139387, + 0.0040231612510979176, + 0.007249223068356514, + 0.005857827607542276, + 0.005281977355480194, + 0.0033840369433164597, + 0.006704946979880333, + 0.006528944242745638, + 0.0014512543566524982, + 0.0019022502237930894, + 0.0029057881329208612, + 0.002083685016259551, + 0.0020985291339457035, + 0.0010390690295025706, + 0.0017049775924533606, + 0.0027079805731773376, + 0.006316795013844967, + 0.005952711217105389, + 0.006572455633431673, + 0.004812563303858042, + 0.004665070213377476, + 0.003537262324243784, + 0.005714463535696268, + 0.008150823414325714, + 0.008667326532304287, + 0.010723029263317585, + 0.04182123392820358, + 0.00905525777488947, + 0.007510773371905088, + 0.006425624713301659, + 0.00801075715571642, + 0.04065581038594246, + 0.016465039923787117, + 0.014798159711062908, + 0.034099239856004715, + 0.01534229889512062, + 0.0442626029253006, + 0.01860782317817211, + 0.030580181628465652, + 0.02872410975396633, + 0.010853350162506104, + 0.007896705530583858, + 0.01850508712232113, + 0.011165771633386612, + 0.007226257119327784, + 0.010689418762922287, + 0.03169211372733116, + 0.047005705535411835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.0009277330245822668 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798814833164215 + ], + "opposite_logprob_tokens": [ + -1.8048814535140991 + ], + "logprob_reward": -0.1798814833164215, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798814833164215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20967364311218262, + 0.055702053010463715, + 0.04900660365819931, + 0.12625223398208618, + 0.04955471679568291, + 0.04424365237355232, + 0.08320645242929459, + 0.05848105996847153, + 0.17816577851772308, + 0.14571379125118256 + ], + "credit_map": [ + 0.012514477595686913, + 0.021991807967424393, + 0.021834976971149445, + 0.02268400602042675, + 0.021350113674998283, + 0.03969722241163254, + 0.03145679458975792, + 0.038144245743751526, + 0.002945488318800926, + 0.006173432804644108, + 0.007140487432479858, + 0.006069821771234274, + 0.005130921956151724, + 0.00742743955925107, + 0.010717067867517471, + 0.010097396560013294, + 0.0026466671843081713, + 0.00703383469954133, + 0.007436767220497131, + 0.0065569500438869, + 0.00747848954051733, + 0.004222686868160963, + 0.008360007777810097, + 0.005271205212920904, + 0.006601002067327499, + 0.012832438573241234, + 0.0364261232316494, + 0.036428261548280716, + 0.011625044979155064, + 0.006031839642673731, + 0.00901913084089756, + 0.007288400549441576, + 0.0035786444786936045, + 0.003947493154555559, + 0.008653095923364162, + 0.008446227759122849, + 0.004114591982215643, + 0.003899619448930025, + 0.00805751420557499, + 0.00885753147304058, + 0.002050012582913041, + 0.005269522313028574, + 0.0034689088352024555, + 0.011755082756280899, + 0.0031728071626275778, + 0.0028519697953015566, + 0.00401656236499548, + 0.011658790521323681, + 0.011560682207345963, + 0.010004628449678421, + 0.011269022710621357, + 0.010011903010308743, + 0.006742189172655344, + 0.00840508658438921, + 0.009111874736845493, + 0.016101069748401642, + 0.007347264792770147, + 0.008723244071006775, + 0.0077605098485946655, + 0.0065268464386463165, + 0.006107962690293789, + 0.005999726243317127, + 0.006230611354112625, + 0.009784895926713943, + 0.01343407854437828, + 0.01577833853662014, + 0.029566267505288124, + 0.0174210537225008, + 0.03827616199851036, + 0.01716095395386219, + 0.02767302095890045, + 0.018855905160307884, + 0.013387220911681652, + 0.009868817403912544, + 0.027617719024419785, + 0.015697861090302467, + 0.017570167779922485, + 0.0156965721398592, + 0.021016113460063934, + 0.024859337136149406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.1798814833164215 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3133974075317383 + ], + "opposite_logprob_tokens": [ + -1.3133974075317383 + ], + "logprob_reward": -0.3133974075317383, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3133974075317383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08254139125347137, + 0.041062019765377045, + 0.03260176628828049, + 0.199921116232872, + 0.04223004728555679, + 0.026880323886871338, + 0.04353541508316994, + 0.09479231387376785, + 0.2633264362812042, + 0.17310915887355804 + ], + "credit_map": [ + 0.00783106591552496, + 0.011144145391881466, + 0.013574776239693165, + 0.013400777243077755, + 0.0078769875690341, + 0.00906143058091402, + 0.010585845448076725, + 0.009066363796591759, + 0.0023220540024340153, + 0.004618157632648945, + 0.004673432093113661, + 0.004249163903295994, + 0.00442025950178504, + 0.005667864345014095, + 0.007282016333192587, + 0.007829077541828156, + 0.0021517735440284014, + 0.006453872658312321, + 0.0081239715218544, + 0.006102022714912891, + 0.002309027360752225, + 0.0016631592297926545, + 0.0028395839035511017, + 0.0029583550058305264, + 0.010324579663574696, + 0.020088424906134605, + 0.05190212279558182, + 0.05008375272154808, + 0.029519928619265556, + 0.008686901070177555, + 0.014827894978225231, + 0.01448751613497734, + 0.003858471056446433, + 0.003852317575365305, + 0.007389839272946119, + 0.008735007606446743, + 0.003389182034879923, + 0.0030185093637555838, + 0.007554870098829269, + 0.004431850742548704, + 0.00225415569730103, + 0.002513871993869543, + 0.003292504698038101, + 0.005597563926130533, + 0.0024707582779228687, + 0.0015150661347433925, + 0.005047192331403494, + 0.00418921047821641, + 0.003826737869530916, + 0.0048802499659359455, + 0.006668472662568092, + 0.004535290878266096, + 0.004087630193680525, + 0.003786642337217927, + 0.006128122564405203, + 0.009622267447412014, + 0.012209950014948845, + 0.014277355745434761, + 0.01165408082306385, + 0.011809428222477436, + 0.009565100073814392, + 0.008968737907707691, + 0.00966031476855278, + 0.016647346317768097, + 0.022945985198020935, + 0.02295655757188797, + 0.05471040681004524, + 0.027882473543286324, + 0.031319256871938705, + 0.03289338946342468, + 0.03594543784856796, + 0.034672923386096954, + 0.009659615345299244, + 0.007854697294533253, + 0.016950573772192, + 0.01123750302940607, + 0.010138074867427349, + 0.01173844188451767, + 0.056434109807014465, + 0.04909614473581314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.3133974075317383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot sweeps across the highly polished wooden lane surface, playing with a gleaming bowling ball. The ball, a vibrant candy-apple red sphere with a perfect mirror-like finish, spins rapidly on its axis, its rapid motion creating motion blur and streaking distorted reflections of the bright overhead fluorescent lights across its surface. It rushes down the varnished maple planks, which themselves shimmer with a high-gloss finish, casting a secondary, elongated reflection of a red blur. The dark, deep gutters on either side create a stark contrast, focusing the viewer's attention on the pristine white pins arranged in a perfect triangle in the distance. The ball moves with swift and decisive force, gliding frictionlessly along its path in a powerful, straight trajectory, creating a tense, focused atmosphere of pure kinetic energy and imminent impact under the cool, even fluorescent light.", + "seed": 6991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_1/index_4132_seed_6991_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": -0.27160219091456383, + "vlm_logprob": -0.27160219091456383, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.27160219091456383, + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23276035487651825 + ], + "opposite_logprob_tokens": [ + -1.9827603101730347 + ], + "logprob_reward": -0.23276035487651825, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.23276035487651825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11147639900445938, + 0.031189413741230965, + 0.024193797260522842, + 0.04668142646551132, + 0.530174195766449, + 0.14898820221424103, + 0.04052020236849785, + 0.015530528500676155, + 0.0351434089243412, + 0.016102349385619164 + ], + "credit_map": [ + 0.011943979188799858, + 0.01144800242036581, + 0.019699692726135254, + 0.01833610236644745, + 0.012993969023227692, + 0.010788439773023129, + 0.0159648135304451, + 0.010301407426595688, + 0.0038630240596830845, + 0.003653433173894882, + 0.003959184046834707, + 0.0036511614453047514, + 0.0038966480642557144, + 0.003933236468583345, + 0.003717828541994095, + 0.004514899104833603, + 0.0029814457520842552, + 0.002232113154605031, + 0.0026508914306759834, + 0.004996969364583492, + 0.0023295520804822445, + 0.003090736921876669, + 0.002938902238383889, + 0.0029731879476457834, + 0.004397232551127672, + 0.006525580305606127, + 0.004615326412022114, + 0.010469228029251099, + 0.004388581030070782, + 0.005807902198284864, + 0.005938678979873657, + 0.004538894630968571, + 0.055052295327186584, + 0.03798946738243103, + 0.20057238638401031, + 0.12387460470199585, + 0.02642321027815342, + 0.030715981498360634, + 0.03515114262700081, + 0.02039511315524578, + 0.02813873253762722, + 0.010518092662096024, + 0.039590928703546524, + 0.019111894071102142, + 0.01730945147573948, + 0.015687167644500732, + 0.009135695174336433, + 0.009496252983808517, + 0.0026978114619851112, + 0.002302823355421424, + 0.00743586290627718, + 0.004883539397269487, + 0.0034186746925115585, + 0.006871967110782862, + 0.005063148681074381, + 0.007846374996006489, + 0.0019172425381839275, + 0.0019726676400750875, + 0.001540499273687601, + 0.00236148783005774, + 0.0016201365506276488, + 0.002258636523038149, + 0.0020992467179894447, + 0.0017606116598472, + 0.005886318162083626, + 0.005342224147170782, + 0.0032115757931023836, + 0.004712678026407957, + 0.0033967983908951283, + 0.003441612236201763, + 0.005036273039877415, + 0.004115928430110216, + 0.0015660561621189117, + 0.0022820387966930866, + 0.001917918212711811, + 0.0025697594974189997, + 0.0029633415397256613, + 0.0012419742997735739, + 0.0014989596093073487, + 0.002062300918623805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.23276035487651825 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486451208591461 + ], + "opposite_logprob_tokens": [ + -1.2236450910568237 + ], + "logprob_reward": -0.3486451208591461, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486451208591461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018948372453451157, + 0.008895969949662685, + 0.06185828894376755, + 0.030320242047309875, + 0.7611486911773682, + 0.07151362299919128, + 0.018863016739487648, + 0.004567746538668871, + 0.008568324148654938, + 0.015315700322389603 + ], + "credit_map": [ + 0.0019420036114752293, + 0.0017903742846101522, + 0.002892028307542205, + 0.0036972451489418745, + 0.0022664235439151525, + 0.0020218081772327423, + 0.0028278047684580088, + 0.0015106851933524013, + 0.000991548877209425, + 0.0010653771460056305, + 0.0009614955051802099, + 0.00099507428240031, + 0.0014501046389341354, + 0.0012584296055138111, + 0.0012411314528435469, + 0.0009328093728981912, + 0.004218215588480234, + 0.0027737803757190704, + 0.013900335878133774, + 0.004587051924318075, + 0.0026129435282200575, + 0.010648660361766815, + 0.0028220394160598516, + 0.020295266062021255, + 0.0035715552512556314, + 0.00385288312099874, + 0.002904332708567381, + 0.0056782253086566925, + 0.004890474956482649, + 0.00295824883505702, + 0.004109438043087721, + 0.0023550866171717644, + 0.0828007236123085, + 0.054537802934646606, + 0.29316040873527527, + 0.1852760910987854, + 0.03420690819621086, + 0.03696818649768829, + 0.04849203675985336, + 0.025706587359309196, + 0.004287390504032373, + 0.00393279641866684, + 0.024004822596907616, + 0.011203404515981674, + 0.011127825826406479, + 0.007156106643378735, + 0.004718459211289883, + 0.005082825198769569, + 0.0020169534254819155, + 0.0011201162124052644, + 0.0035747848451137543, + 0.0017408833373337984, + 0.001338503323495388, + 0.0037133304867893457, + 0.0016647225711494684, + 0.0036937242839485407, + 0.0004919023485854268, + 0.0005277199088595808, + 0.00045316177420318127, + 0.0007512154406867921, + 0.0003798629913944751, + 0.0007555072661489248, + 0.0004588362935464829, + 0.0007495409809052944, + 0.0011541094863787293, + 0.0010125739499926567, + 0.0007976202177815139, + 0.0010183145059272647, + 0.0008237165748141706, + 0.001375454245135188, + 0.0010058778570964932, + 0.0013806584756821394, + 0.0019112558802589774, + 0.0020734602585434914, + 0.0023120774421840906, + 0.003220735816285014, + 0.0019179488299414515, + 0.0012951863463968039, + 0.0012929189251735806, + 0.0012921176385134459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.3486451208591461 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009380945703014731 + ], + "opposite_logprob_tokens": [ + -7.0009379386901855 + ], + "logprob_reward": -0.0009380945703014731, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009380945703014731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0989718958735466, + 0.06915758550167084, + 0.20542097091674805, + 0.06931587308645248, + 0.3787193298339844, + 0.08331358432769775, + 0.04681480675935745, + 0.011721281334757805, + 0.016261758282780647, + 0.020302927121520042 + ], + "credit_map": [ + 0.011151788756251335, + 0.011397571302950382, + 0.016002057120203972, + 0.016720497980713844, + 0.01175627764314413, + 0.013031801208853722, + 0.011068224906921387, + 0.00784367322921753, + 0.007288838271051645, + 0.008534105494618416, + 0.0073869600892066956, + 0.010878082364797592, + 0.01044463086873293, + 0.008279887959361076, + 0.008913268335163593, + 0.007431809324771166, + 0.013102141208946705, + 0.009559094905853271, + 0.04423965513706207, + 0.016277452930808067, + 0.012171714566648006, + 0.03556353226304054, + 0.01229977235198021, + 0.06220760941505432, + 0.007527448236942291, + 0.010577312670648098, + 0.005975392647087574, + 0.012243247590959072, + 0.011439808644354343, + 0.007023345213383436, + 0.010163289494812489, + 0.004366032313555479, + 0.0403190441429615, + 0.027976008132100105, + 0.1413022130727768, + 0.09089785069227219, + 0.018139846622943878, + 0.019278813153505325, + 0.026618655771017075, + 0.014186887070536613, + 0.011588504537940025, + 0.0051542650908231735, + 0.03203257545828819, + 0.014554893597960472, + 0.005898701958358288, + 0.004864651244133711, + 0.0060926442965865135, + 0.003127348143607378, + 0.0030054801609367132, + 0.002775688888505101, + 0.007093367632478476, + 0.009888017550110817, + 0.0035449673887342215, + 0.005972422193735838, + 0.0076040299609303474, + 0.006930836010724306, + 0.0014316225424408913, + 0.001296629081480205, + 0.0010632664198055863, + 0.0014677427243441343, + 0.0011963126016780734, + 0.0019469800172373652, + 0.0012718746438622475, + 0.0020468533039093018, + 0.002235074993222952, + 0.0022201831452548504, + 0.001549887820146978, + 0.0021351408213377, + 0.002038132632151246, + 0.0021560420282185078, + 0.0017930861795321107, + 0.0021342101972550154, + 0.0034605227410793304, + 0.0021148782689124346, + 0.0030549573712050915, + 0.004514819011092186, + 0.002529909834265709, + 0.001397266169078648, + 0.001701488858088851, + 0.0015290846349671483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0009380945703014731 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284159243106842 + ], + "opposite_logprob_tokens": [ + -2.0178415775299072 + ], + "logprob_reward": -0.14284159243106842, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284159243106842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03203600272536278, + 0.019584594294428825, + 0.08671466261148453, + 0.030278220772743225, + 0.5958796739578247, + 0.13082322478294373, + 0.03851843997836113, + 0.013691033236682415, + 0.016301704570651054, + 0.036172494292259216 + ], + "credit_map": [ + 0.0041106767021119595, + 0.003827374894171953, + 0.004769288003444672, + 0.004954005591571331, + 0.004201632924377918, + 0.00342367822304368, + 0.004021193366497755, + 0.0027281520888209343, + 0.0017118219984695315, + 0.0020047796424478292, + 0.0023948545567691326, + 0.002365036169067025, + 0.003003740916028619, + 0.0025930500123649836, + 0.0035210144706070423, + 0.0019902957137674093, + 0.006407742854207754, + 0.004412288777530193, + 0.019115939736366272, + 0.006723351310938597, + 0.003935396671295166, + 0.01471875049173832, + 0.004195383284240961, + 0.0272058118134737, + 0.0033958640415221453, + 0.003842683508992195, + 0.0028015864081680775, + 0.00620918907225132, + 0.004559800960123539, + 0.003019638592377305, + 0.0041541121900081635, + 0.0022953455336391926, + 0.06624779105186462, + 0.04061912000179291, + 0.23256610333919525, + 0.144198939204216, + 0.02690771594643593, + 0.027778517454862595, + 0.0378701314330101, + 0.019691327586770058, + 0.007879921235144138, + 0.007646678946912289, + 0.04479483515024185, + 0.021361667662858963, + 0.02547070011496544, + 0.008477934636175632, + 0.008577624335885048, + 0.006613869220018387, + 0.0028650150634348392, + 0.0022384249605238438, + 0.005620378535240889, + 0.007963169366121292, + 0.0026597806718200445, + 0.005425488576292992, + 0.006762244738638401, + 0.004983940627425909, + 0.0015015997923910618, + 0.0016357809072360396, + 0.0013076600152999163, + 0.001429716357961297, + 0.0012235479662194848, + 0.0025444752536714077, + 0.0014622227754443884, + 0.0025860308669507504, + 0.0020005342084914446, + 0.00187496782746166, + 0.00145635602530092, + 0.0020790318958461285, + 0.0014614415122196078, + 0.0026386003009974957, + 0.0016046948730945587, + 0.00318607734516263, + 0.003353967797011137, + 0.006379412487149239, + 0.0044112117029726505, + 0.005684243515133858, + 0.007623041048645973, + 0.001955966465175152, + 0.003733596997335553, + 0.0030310535803437233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.14284159243106842 + }, + { + "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\": \"Is the camera performing a tracking shot that sweeps across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant candy-apple red bowling ball with a mirror-like finish visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the bowling ball spinning rapidly on its axis while moving down the lane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are white bowling pins arranged in a perfect triangle visible in the distance?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328257918357849 + ], + "opposite_logprob_tokens": [ + -0.7578257918357849 + ], + "logprob_reward": -0.6328257918357849, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328257918357849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04964901879429817, + 0.09158995002508163, + 0.2777899503707886, + 0.018992885947227478, + 0.3919505774974823, + 0.061882779002189636, + 0.017965707927942276, + 0.015712324529886246, + 0.02354394644498825, + 0.05092289671301842 + ], + "credit_map": [ + 0.0060258107259869576, + 0.004815541673451662, + 0.006613763980567455, + 0.009080830961465836, + 0.005398016422986984, + 0.006527421064674854, + 0.006561534479260445, + 0.004626098554581404, + 0.009553865529596806, + 0.008443959057331085, + 0.014438479207456112, + 0.015913808718323708, + 0.006840186193585396, + 0.015117213129997253, + 0.010248323902487755, + 0.011034096591174603, + 0.018555887043476105, + 0.010560039430856705, + 0.0661197081208229, + 0.01750131882727146, + 0.01332672405987978, + 0.04408279433846474, + 0.014240233227610588, + 0.09340320527553558, + 0.002460208721458912, + 0.002643959829583764, + 0.002212862018495798, + 0.0036485849414020777, + 0.0016539179487153888, + 0.0025246606674045324, + 0.0024147999938577414, + 0.0014338911278173327, + 0.042581621557474136, + 0.02681259624660015, + 0.15237930417060852, + 0.09423385560512543, + 0.017540238797664642, + 0.02048065885901451, + 0.025129100307822227, + 0.012793182395398617, + 0.007517909165471792, + 0.003046763362362981, + 0.02126944437623024, + 0.009674468077719212, + 0.007060405798256397, + 0.004949323832988739, + 0.0036671310663223267, + 0.004697327967733145, + 0.0015231441939249635, + 0.001191379502415657, + 0.003877747803926468, + 0.0018368796445429325, + 0.0010025082156062126, + 0.0029395066667348146, + 0.0021513174287974834, + 0.0034432215616106987, + 0.0015588609967380762, + 0.0025834382977336645, + 0.0014812133740633726, + 0.0014220327138900757, + 0.0011945284204557538, + 0.002810330595821142, + 0.0021096724085509777, + 0.002552247140556574, + 0.002375583630055189, + 0.0023472311440855265, + 0.0023493466433137655, + 0.002691807458177209, + 0.0021588962990790606, + 0.004193114582449198, + 0.002379870740696788, + 0.005048095248639584, + 0.009694894775748253, + 0.0072657219134271145, + 0.007216585334390402, + 0.008718929253518581, + 0.007030515465885401, + 0.003374719526618719, + 0.003405399853363633, + 0.004216126166284084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.6328257918357849 + } + ], + "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/0000167/group_10/index_tensor([6387, 6387])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_tensor([6387, 6387])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7008bb847484d3f64a4621929c9cb1cfec90d22f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_tensor([6387, 6387])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly panning drone shot captures the majestic sight of a vintage steam locomotive steadily and powerfully traversing a vast field. The entire scene takes on the quality of a dynamic oil painting, characterized by thick brushstrokes and vibrant colors. From above, the landscape resembles a vibrant canvas of golden wheat fields, emerald green meadows, and ochre soil, bisected by two parallel railroad tracks like a silver ribbon. A glossy black locomotive pulls several crimson and cream carriages along the straight track. Thick white steam billows from its chimney, leaving a long, gradually dissipating trail in the air, its form unmistakable, as if painted with a thick brush. The warm afternoon sun slants down, casting the train's long shadow across the field and reflecting a fluid sheen off its metal body. The colors are saturated and richly layered, and the brushstrokes are clearly visible in the textures of the fields and the drifting steam, creating an atmosphere of tranquility that is both grand and full of nostalgic poetry.", + "seed": 14671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14671_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": -0.09234763870481402, + "vlm_logprob": -0.09234763870481402, + "vlm_yes_no_margin": 4.675000011920929, + "vlm_token_credit": -0.09234763870481402, + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4400288462638855 + ], + "opposite_logprob_tokens": [ + -1.0650289058685303 + ], + "logprob_reward": -0.4400288462638855, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4400288462638855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15958237648010254, + 0.18107609450817108, + 0.17916132509708405, + 0.06747394055128098, + 0.04154828190803528, + 0.11597000062465668, + 0.10140170902013779, + 0.052780456840991974, + 0.07164518535137177, + 0.029360653832554817 + ], + "credit_map": [ + 0.02454352006316185, + 0.017778674140572548, + 0.01650995947420597, + 0.01605350337922573, + 0.016324587166309357, + 0.02217152528464794, + 0.016703851521015167, + 0.029496753588318825, + 0.015391149558126926, + 0.021011272445321083, + 0.024783039465546608, + 0.03567606583237648, + 0.015349478460848331, + 0.022622212767601013, + 0.022191345691680908, + 0.02405153028666973, + 0.02897913008928299, + 0.019834013655781746, + 0.02707490138709545, + 0.030715743079781532, + 0.012521828524768353, + 0.01835745759308338, + 0.021197045221924782, + 0.020481178537011147, + 0.00932298693805933, + 0.009361924603581429, + 0.008574790321290493, + 0.006325258873403072, + 0.012686964124441147, + 0.009105752222239971, + 0.006164013873785734, + 0.005932244006544352, + 0.0057336692698299885, + 0.007266539614647627, + 0.006482438649982214, + 0.005940191447734833, + 0.005800630431622267, + 0.004499303176999092, + 0.002910891780629754, + 0.0029146154411137104, + 0.01037237886339426, + 0.04028061032295227, + 0.005594147834926844, + 0.006048143841326237, + 0.02646975964307785, + 0.012821191921830177, + 0.006171720568090677, + 0.008212038315832615, + 0.012105204164981842, + 0.009877828881144524, + 0.015156419016420841, + 0.01032189466059208, + 0.012303746305406094, + 0.01485682837665081, + 0.01426553912460804, + 0.012514237314462662, + 0.013588204979896545, + 0.0046773748472332954, + 0.0056281606666743755, + 0.006306998897343874, + 0.004647811874747276, + 0.004642662592232227, + 0.003901019459590316, + 0.009388218633830547, + 0.013157065957784653, + 0.009945734404027462, + 0.00687151774764061, + 0.004290775395929813, + 0.008645860478281975, + 0.013875525444746017, + 0.006213505286723375, + 0.008645196445286274, + 0.0065320078283548355, + 0.0041624316945672035, + 0.002767767757177353, + 0.0019601930398494005, + 0.0043306536972522736, + 0.004978364333510399, + 0.00251290132291615, + 0.002116332994773984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.4400288462638855 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004660695791244507 + ], + "opposite_logprob_tokens": [ + -5.379660606384277 + ], + "logprob_reward": -0.004660695791244507, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004660695791244507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2961445152759552, + 0.2316315770149231, + 0.09992068260908127, + 0.062218982726335526, + 0.021790994331240654, + 0.08233527094125748, + 0.048063866794109344, + 0.09109031409025192, + 0.016510579735040665, + 0.050293128937482834 + ], + "credit_map": [ + 0.02730519510805607, + 0.023004787042737007, + 0.03176206350326538, + 0.03425374627113342, + 0.026730624958872795, + 0.04302871599793434, + 0.03118331916630268, + 0.0788760855793953, + 0.0208606980741024, + 0.01923520304262638, + 0.03501676395535469, + 0.057180941104888916, + 0.015578605234622955, + 0.028202177956700325, + 0.02606385387480259, + 0.029493365436792374, + 0.014762302860617638, + 0.011197465471923351, + 0.012950863689184189, + 0.020221073180437088, + 0.006077466066926718, + 0.011182677000761032, + 0.012711464427411556, + 0.010817382484674454, + 0.008708861656486988, + 0.009577625431120396, + 0.008032959885895252, + 0.005435593891888857, + 0.011461395770311356, + 0.008580627851188183, + 0.005168356467038393, + 0.005253568757325411, + 0.0025171968154609203, + 0.004827187862247229, + 0.0030163568444550037, + 0.0031482644844800234, + 0.0024038394913077354, + 0.001961646368727088, + 0.0022022572811692953, + 0.001714246696792543, + 0.0074376375414431095, + 0.017872458323836327, + 0.01740170456469059, + 0.011373773217201233, + 0.010042189620435238, + 0.0070627485401928425, + 0.0053892615251243114, + 0.005755500867962837, + 0.006298871245235205, + 0.0046241143718361855, + 0.007287490181624889, + 0.005676703527569771, + 0.006220349110662937, + 0.005000918172299862, + 0.007320386357605457, + 0.005635038949549198, + 0.02715224213898182, + 0.010073993355035782, + 0.010313136503100395, + 0.008611645549535751, + 0.0075351884588599205, + 0.005978865083307028, + 0.005654910113662481, + 0.01577034406363964, + 0.0016003758646547794, + 0.0017308383248746395, + 0.0021821411792188883, + 0.0022525866515934467, + 0.0022971576545387506, + 0.003262259066104889, + 0.0014217088464647532, + 0.0017635165713727474, + 0.014676062390208244, + 0.010228327475488186, + 0.004319127183407545, + 0.004189133644104004, + 0.0058876872062683105, + 0.004376863595098257, + 0.003250776557251811, + 0.00336515624076128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004660695791244507 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197180047631264 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197180047631264, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197180047631264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553909033536911, + 0.20018765330314636, + 0.28397542238235474, + 0.13027961552143097, + 0.028805579990148544, + 0.05178197845816612, + 0.040409430861473083, + 0.06414089351892471, + 0.022416852414608, + 0.02261166088283062 + ], + "credit_map": [ + 0.02515779435634613, + 0.013537844642996788, + 0.015750164166092873, + 0.01689622923731804, + 0.017104467377066612, + 0.021760068833827972, + 0.014450709335505962, + 0.030733618885278702, + 0.018337862566113472, + 0.02046886272728443, + 0.025203155353665352, + 0.0429774634540081, + 0.01779945380985737, + 0.02599352039396763, + 0.019043834879994392, + 0.030363498255610466, + 0.03752591088414192, + 0.026936909183859825, + 0.04951029270887375, + 0.05347588285803795, + 0.019117508083581924, + 0.029455283656716347, + 0.03686574846506119, + 0.031087888404726982, + 0.021835504099726677, + 0.021295087411999702, + 0.013290663249790668, + 0.010089599527418613, + 0.02655177004635334, + 0.01578514091670513, + 0.009506640955805779, + 0.011925188824534416, + 0.003838334698230028, + 0.0048918710090219975, + 0.003092458238825202, + 0.00411009369418025, + 0.004476901143789291, + 0.0033792334143072367, + 0.0029475705232471228, + 0.0020691165700554848, + 0.0046770269982516766, + 0.010210581123828888, + 0.01181240938603878, + 0.008092156611382961, + 0.005354593507945538, + 0.004589426331222057, + 0.003209754591807723, + 0.003836028976365924, + 0.005247724708169699, + 0.004150030668824911, + 0.00668372493237257, + 0.005062193144112825, + 0.005683769937604666, + 0.004545657429844141, + 0.004969553556293249, + 0.004066772293299437, + 0.009041390381753445, + 0.009735560044646263, + 0.005096863489598036, + 0.004378028679639101, + 0.008140955120325089, + 0.013158464804291725, + 0.006263154558837414, + 0.00832646619528532, + 0.002629467286169529, + 0.0025265386793762445, + 0.002767554484307766, + 0.0018238041084259748, + 0.003548073349520564, + 0.003044448560103774, + 0.003059670329093933, + 0.003017294919118285, + 0.00664793886244297, + 0.003309519961476326, + 0.0017525925068184733, + 0.0012724699918180704, + 0.0028153802268207073, + 0.0033657574094831944, + 0.0015252700541168451, + 0.0019227316370233893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014197180047631264 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013380631571635604 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013380631571635604, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013380631571635604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1377691626548767, + 0.15986870229244232, + 0.28424859046936035, + 0.07317081838846207, + 0.08637918531894684, + 0.08360408991575241, + 0.07585843652486801, + 0.04520563781261444, + 0.03476633504033089, + 0.019128968939185143 + ], + "credit_map": [ + 0.022784747183322906, + 0.013071676716208458, + 0.014125542715191841, + 0.015759969130158424, + 0.014454945921897888, + 0.018815848976373672, + 0.013768934644758701, + 0.024987488985061646, + 0.015502947382628918, + 0.013776114210486412, + 0.023409195244312286, + 0.03701222687959671, + 0.00972270593047142, + 0.02070915699005127, + 0.017480922862887383, + 0.022255441173911095, + 0.03938213363289833, + 0.028967617079615593, + 0.04737699404358864, + 0.05219031497836113, + 0.019847046583890915, + 0.03044780157506466, + 0.03404003009200096, + 0.03199666738510132, + 0.010989556089043617, + 0.009685304947197437, + 0.009518851526081562, + 0.008464610204100609, + 0.012382623739540577, + 0.007170278113335371, + 0.0072937654331326485, + 0.007665832992643118, + 0.01236022263765335, + 0.012561412528157234, + 0.01270258892327547, + 0.013232355006039143, + 0.011032863520085812, + 0.009009234607219696, + 0.009126919321715832, + 0.006353598553687334, + 0.006752137094736099, + 0.027882404625415802, + 0.00603138655424118, + 0.004814523737877607, + 0.018474511802196503, + 0.00954798050224781, + 0.004844063427299261, + 0.005257078446447849, + 0.009882817044854164, + 0.008051070384681225, + 0.012315132655203342, + 0.007483238819986582, + 0.007383512798696756, + 0.010059601627290249, + 0.01090491097420454, + 0.009778155013918877, + 0.009276055730879307, + 0.006595320999622345, + 0.005514051765203476, + 0.005153358448296785, + 0.0048684487119317055, + 0.004548836499452591, + 0.004334357567131519, + 0.0049152071587741375, + 0.005641257390379906, + 0.004217865876853466, + 0.004075496457517147, + 0.002661748556420207, + 0.003942109644412994, + 0.0062665739096701145, + 0.00355633907020092, + 0.00440494529902935, + 0.005108969286084175, + 0.0035475222393870354, + 0.0015064696781337261, + 0.001784497406333685, + 0.0021802843548357487, + 0.0022488515824079514, + 0.0013073703739792109, + 0.0014450039016082883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0013380631571635604 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001513408264145255 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.001513408264145255, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001513408264145255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12789134681224823, + 0.32916536927223206, + 0.22058162093162537, + 0.06748242676258087, + 0.033481478691101074, + 0.048852648586034775, + 0.04923008009791374, + 0.07884901016950607, + 0.017699597403407097, + 0.02676638588309288 + ], + "credit_map": [ + 0.0160305704921484, + 0.026118213310837746, + 0.012327381409704685, + 0.01005223672837019, + 0.02077757939696312, + 0.017594970762729645, + 0.011424710042774677, + 0.01356569491326809, + 0.02944197878241539, + 0.046056706458330154, + 0.036783043295145035, + 0.055801112204790115, + 0.042823754251003265, + 0.04382196441292763, + 0.026405151933431625, + 0.048031654208898544, + 0.033303894102573395, + 0.022665714845061302, + 0.03611551597714424, + 0.040323346853256226, + 0.014467663131654263, + 0.022751573473215103, + 0.02704831399023533, + 0.02390560507774353, + 0.009218367747962475, + 0.010257603600621223, + 0.008955180644989014, + 0.006342773791402578, + 0.01194539200514555, + 0.00897900853306055, + 0.005818833131343126, + 0.005965269636362791, + 0.0047141388058662415, + 0.004317067563533783, + 0.004507355857640505, + 0.004323213826864958, + 0.00431471224874258, + 0.004182049538940191, + 0.004617426544427872, + 0.002505513373762369, + 0.004275277256965637, + 0.013290434144437313, + 0.007326118648052216, + 0.0052846320904791355, + 0.007979229092597961, + 0.004731642547994852, + 0.002767179161310196, + 0.0031981347128748894, + 0.005500764586031437, + 0.004716743249446154, + 0.007514606695622206, + 0.0054419199004769325, + 0.005814829375594854, + 0.0066089206375181675, + 0.007639308925718069, + 0.005992986727505922, + 0.02671484835445881, + 0.006817227229475975, + 0.005978160537779331, + 0.005576615687459707, + 0.0055909547954797745, + 0.004754884634166956, + 0.0050034066662192345, + 0.018412912264466286, + 0.0022792278323322535, + 0.002420103410258889, + 0.0020621188450604677, + 0.0020084797870367765, + 0.0025120906066149473, + 0.0027301518712192774, + 0.0019212150946259499, + 0.0017662107711657882, + 0.007758659310638905, + 0.004886722192168236, + 0.0015975869027897716, + 0.002056245459243655, + 0.0026509554591029882, + 0.004369499161839485, + 0.0016257315874099731, + 0.0018209881382063031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.001513408264145255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly panning drone shot captures the majestic sight of a vintage steam locomotive steadily and powerfully traversing a vast field. The entire scene takes on the quality of a dynamic oil painting, characterized by thick brushstrokes and vibrant colors. From above, the landscape resembles a vibrant canvas of golden wheat fields, emerald green meadows, and ochre soil, bisected by two parallel railroad tracks like a silver ribbon. A glossy black locomotive pulls several crimson and cream carriages along the straight track. Thick white steam billows from its chimney, leaving a long, gradually dissipating trail in the air, its form unmistakable, as if painted with a thick brush. The warm afternoon sun slants down, casting the train's long shadow across the field and reflecting a fluid sheen off its metal body. The colors are saturated and richly layered, and the brushstrokes are clearly visible in the textures of the fields and the drifting steam, creating an atmosphere of tranquility that is both grand and full of nostalgic poetry.", + "seed": 14672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14672_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": -0.11407522009685636, + "vlm_logprob": -0.11407522009685636, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.11407522009685636, + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4966925382614136 + ], + "opposite_logprob_tokens": [ + -0.9966925382614136 + ], + "logprob_reward": -0.4966925382614136, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4966925382614136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1460370272397995, + 0.09653603285551071, + 0.08004512637853622, + 0.29350095987319946, + 0.04964379593729973, + 0.03721793740987778, + 0.11082059144973755, + 0.04087290167808533, + 0.11734411865472794, + 0.027981441468000412 + ], + "credit_map": [ + 0.017897039651870728, + 0.021731503307819366, + 0.01709696277976036, + 0.01789644919335842, + 0.01855587773025036, + 0.02028593048453331, + 0.015198195353150368, + 0.017375081777572632, + 0.00968144554644823, + 0.01017851009964943, + 0.011241032741963863, + 0.010648053139448166, + 0.0100051024928689, + 0.014971193857491016, + 0.017751209437847137, + 0.012059499509632587, + 0.010049041360616684, + 0.008172564208507538, + 0.009920290671288967, + 0.013489402830600739, + 0.00966800656169653, + 0.008318864740431309, + 0.011068037711083889, + 0.009358922950923443, + 0.014230438508093357, + 0.007618216332048178, + 0.011355169117450714, + 0.016646958887577057, + 0.012322977185249329, + 0.07937654107809067, + 0.029621340334415436, + 0.12232936918735504, + 0.005749314092099667, + 0.006092652212828398, + 0.004929849412292242, + 0.005656485445797443, + 0.007021908648312092, + 0.004528414458036423, + 0.0076126414351165295, + 0.00805253628641367, + 0.004736512433737516, + 0.003651418723165989, + 0.00442085787653923, + 0.004573376849293709, + 0.0041038342751562595, + 0.004403280559927225, + 0.004769300110638142, + 0.006559360772371292, + 0.010096738114953041, + 0.012261106632649899, + 0.013327393680810928, + 0.014982792548835278, + 0.010484889149665833, + 0.012208500877022743, + 0.02066321298480034, + 0.016795964911580086, + 0.004363626707345247, + 0.003041009185835719, + 0.0027075696270912886, + 0.003684260416775942, + 0.004310554824769497, + 0.007132499013096094, + 0.005114011466503143, + 0.010519376955926418, + 0.01437456626445055, + 0.0165347121655941, + 0.00871405377984047, + 0.009269267320632935, + 0.025085395202040672, + 0.012010308913886547, + 0.01114452350884676, + 0.020211314782500267, + 0.003859397489577532, + 0.005394196603447199, + 0.0025121537037193775, + 0.003456051927059889, + 0.004048212431371212, + 0.0027884910814464092, + 0.0026044356636703014, + 0.0033185062929987907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.4966925382614136 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368139311671257 + ], + "opposite_logprob_tokens": [ + -3.6513681411743164 + ], + "logprob_reward": -0.026368139311671257, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368139311671257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21868892014026642, + 0.05049458518624306, + 0.0351749062538147, + 0.05601175129413605, + 0.1670442670583725, + 0.059128303080797195, + 0.3165040612220764, + 0.032918788492679596, + 0.03778330609202385, + 0.026251133531332016 + ], + "credit_map": [ + 0.02449055016040802, + 0.03007666766643524, + 0.03384694084525108, + 0.026994211599230766, + 0.022503593936562538, + 0.022457608953118324, + 0.030759451910853386, + 0.027559906244277954, + 0.003897286020219326, + 0.006657675374299288, + 0.004281758330762386, + 0.004596678540110588, + 0.0035918294452130795, + 0.010908026248216629, + 0.012464510276913643, + 0.004096820019185543, + 0.0047041927464306355, + 0.004801148548722267, + 0.003433345351368189, + 0.003936595283448696, + 0.005048354156315327, + 0.004872546996921301, + 0.0041816094890236855, + 0.004197111818939447, + 0.005281059071421623, + 0.009907894767820835, + 0.00534482067450881, + 0.008419450372457504, + 0.008170828223228455, + 0.005437500774860382, + 0.005522063467651606, + 0.007928134873509407, + 0.013743681833148003, + 0.014666063711047173, + 0.012060950510203838, + 0.017235800623893738, + 0.020060762763023376, + 0.03502042591571808, + 0.03891346603631973, + 0.015343109145760536, + 0.005109669174998999, + 0.005997392348945141, + 0.003983184229582548, + 0.005348019767552614, + 0.006039263214915991, + 0.011790761724114418, + 0.004646322689950466, + 0.01621369458734989, + 0.021132880821824074, + 0.04631353169679642, + 0.022363489493727684, + 0.03444937244057655, + 0.039917953312397, + 0.0464838445186615, + 0.037556033581495285, + 0.0682869479060173, + 0.0035079545341432095, + 0.0032884126994758844, + 0.0038211853243410587, + 0.006239922717213631, + 0.0030887932516634464, + 0.0030554719269275665, + 0.0047895461320877075, + 0.005127498880028725, + 0.005226684268563986, + 0.003999736625701189, + 0.003381552640348673, + 0.0034366874024271965, + 0.005635529290884733, + 0.004019733518362045, + 0.004396991804242134, + 0.0076863933354616165, + 0.0034795464016497135, + 0.004348816350102425, + 0.0020647398196160793, + 0.0033731572329998016, + 0.0032989326864480972, + 0.003972870763391256, + 0.002245725831016898, + 0.003467344678938389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.026368139311671257 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737136282026768 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737136282026768, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737136282026768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3734819293022156, + 0.059420932084321976, + 0.058942269533872604, + 0.0767219066619873, + 0.0779319480061531, + 0.01958523690700531, + 0.10250352323055267, + 0.013937876559793949, + 0.17488282918930054, + 0.04259151220321655 + ], + "credit_map": [ + 0.027374938130378723, + 0.024506239220499992, + 0.03516409173607826, + 0.03249221667647362, + 0.061591725796461105, + 0.08828584849834442, + 0.05137781798839569, + 0.05268905684351921, + 0.005849326495081186, + 0.007912020199000835, + 0.005844762083142996, + 0.00863018911331892, + 0.00721088657155633, + 0.007640967145562172, + 0.008200197480618954, + 0.008132589980959892, + 0.0038802556227892637, + 0.006649106740951538, + 0.006692569237202406, + 0.007656088098883629, + 0.005344144534319639, + 0.007625605445355177, + 0.010205966420471668, + 0.010888536460697651, + 0.005050208419561386, + 0.004580915905535221, + 0.007480695378035307, + 0.006800683215260506, + 0.005135039798915386, + 0.016083505004644394, + 0.007933264598250389, + 0.023657603189349174, + 0.006244870368391275, + 0.006946186535060406, + 0.009614620357751846, + 0.008218596689403057, + 0.009007366374135017, + 0.010409614071249962, + 0.01260451227426529, + 0.014886196702718735, + 0.0023874137550592422, + 0.0021139723248779774, + 0.002398525597527623, + 0.0029457020573318005, + 0.0020537390373647213, + 0.0019960973877459764, + 0.002842287765815854, + 0.002847501775249839, + 0.007848039269447327, + 0.01137140579521656, + 0.009874759241938591, + 0.012234404683113098, + 0.010483402758836746, + 0.015293371863663197, + 0.014582566916942596, + 0.02081558108329773, + 0.0015843502478674054, + 0.001269895932637155, + 0.0013031380949541926, + 0.0018904614262282848, + 0.0013162506511434913, + 0.0020205893088132143, + 0.0018718547653406858, + 0.0026813377626240253, + 0.02031450904905796, + 0.023780060932040215, + 0.01462955679744482, + 0.015792446210980415, + 0.032835453748703, + 0.01686045527458191, + 0.01679513230919838, + 0.033875223249197006, + 0.006024817004799843, + 0.0069855572655797005, + 0.0029004618991166353, + 0.004730842541903257, + 0.004702794831246138, + 0.008449338376522064, + 0.0034454797860234976, + 0.005352222826331854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.006737136282026768 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14880268275737762, + 0.09553933888673782, + 0.07379836589097977, + 0.18594709038734436, + 0.08230852335691452, + 0.07231498509645462, + 0.12002065777778625, + 0.04444779083132744, + 0.1351059228181839, + 0.04171469062566757 + ], + "credit_map": [ + 0.01744721457362175, + 0.013081971555948257, + 0.012717662379145622, + 0.012695587240159512, + 0.02293236553668976, + 0.029771752655506134, + 0.021388093009591103, + 0.018768033012747765, + 0.009811636991798878, + 0.01048043929040432, + 0.010794172063469887, + 0.013478496111929417, + 0.009326844476163387, + 0.012948952615261078, + 0.015562727116048336, + 0.013136063702404499, + 0.007154711522161961, + 0.009185699746012688, + 0.006946098525077105, + 0.010146385990083218, + 0.008318074978888035, + 0.0069760228507220745, + 0.016213400289416313, + 0.008857972919940948, + 0.00829011108726263, + 0.006409718655049801, + 0.007946758531033993, + 0.011447248049080372, + 0.009274067357182503, + 0.04848523810505867, + 0.018840620294213295, + 0.07525333762168884, + 0.007868336513638496, + 0.010495039634406567, + 0.010273752734065056, + 0.008523043245077133, + 0.01253686472773552, + 0.00987866148352623, + 0.012389459647238255, + 0.010343365371227264, + 0.005155510269105434, + 0.004372637718915939, + 0.005871344357728958, + 0.006609438452869654, + 0.004070506896823645, + 0.016544917598366737, + 0.0059874230064451694, + 0.023703204467892647, + 0.009698009118437767, + 0.015297027304768562, + 0.01357105653733015, + 0.018344173207879066, + 0.013284657150506973, + 0.01101877260953188, + 0.021480323746800423, + 0.01732664182782173, + 0.007458607200533152, + 0.0028828182257711887, + 0.0039377580396831036, + 0.005019695498049259, + 0.0034804216120392084, + 0.00810445100069046, + 0.005730756558477879, + 0.007833282463252544, + 0.015990667045116425, + 0.01868251897394657, + 0.012093442492187023, + 0.01214086264371872, + 0.026059070602059364, + 0.012665516696870327, + 0.013037643395364285, + 0.02443620003759861, + 0.005956650245934725, + 0.00794161856174469, + 0.0033097565174102783, + 0.004361963830888271, + 0.005102239549160004, + 0.004232906736433506, + 0.004649398382753134, + 0.006160156335681677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0024903970770537853 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038087889552116394 + ], + "opposite_logprob_tokens": [ + -3.288087844848633 + ], + "logprob_reward": -0.038087889552116394, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038087889552116394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2431524246931076, + 0.1572251170873642, + 0.05990452319383621, + 0.11675357818603516, + 0.0885728970170021, + 0.05835975706577301, + 0.07941843569278717, + 0.04226062819361687, + 0.0525766983628273, + 0.1017758920788765 + ], + "credit_map": [ + 0.0248090997338295, + 0.03540770709514618, + 0.030916422605514526, + 0.029289789497852325, + 0.026698151603341103, + 0.035686708986759186, + 0.029583506286144257, + 0.03076103888452053, + 0.01082692202180624, + 0.014248552732169628, + 0.008473322726786137, + 0.011203254573047161, + 0.011738980188965797, + 0.042471885681152344, + 0.04515925422310829, + 0.013102956116199493, + 0.0065432265400886536, + 0.009029309265315533, + 0.006681048311293125, + 0.00886415783315897, + 0.006921519059687853, + 0.005339846946299076, + 0.00940678920596838, + 0.007118625100702047, + 0.009938530623912811, + 0.00897996500134468, + 0.010599293746054173, + 0.009409399703145027, + 0.008246059529483318, + 0.019873080775141716, + 0.012392675504088402, + 0.037314578890800476, + 0.009664799086749554, + 0.010317479260265827, + 0.007848367094993591, + 0.009011313319206238, + 0.015204445458948612, + 0.008288294076919556, + 0.013055698946118355, + 0.015182502567768097, + 0.0038256500847637653, + 0.0033817037474364042, + 0.003379772650077939, + 0.007298951968550682, + 0.0035224631428718567, + 0.013852503150701523, + 0.004010147415101528, + 0.019088570028543472, + 0.009170245379209518, + 0.009473714977502823, + 0.01254244428128004, + 0.011761599220335484, + 0.008021923713386059, + 0.005505589302629232, + 0.013339298777282238, + 0.009603620506823063, + 0.00827073771506548, + 0.006029949523508549, + 0.0025106898974627256, + 0.004423819947987795, + 0.005464596673846245, + 0.004458948504179716, + 0.00466239033266902, + 0.006439495831727982, + 0.005420483183115721, + 0.00734723312780261, + 0.0050799050368368626, + 0.006168460939079523, + 0.006962825544178486, + 0.004444858990609646, + 0.006635749712586403, + 0.01051718182861805, + 0.015505396761000156, + 0.021348722279071808, + 0.008160675875842571, + 0.01075587049126625, + 0.01691102236509323, + 0.011365839280188084, + 0.007694440893828869, + 0.0100339250639081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.038087889552116394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly panning drone shot captures the majestic sight of a vintage steam locomotive steadily and powerfully traversing a vast field. The entire scene takes on the quality of a dynamic oil painting, characterized by thick brushstrokes and vibrant colors. From above, the landscape resembles a vibrant canvas of golden wheat fields, emerald green meadows, and ochre soil, bisected by two parallel railroad tracks like a silver ribbon. A glossy black locomotive pulls several crimson and cream carriages along the straight track. Thick white steam billows from its chimney, leaving a long, gradually dissipating trail in the air, its form unmistakable, as if painted with a thick brush. The warm afternoon sun slants down, casting the train's long shadow across the field and reflecting a fluid sheen off its metal body. The colors are saturated and richly layered, and the brushstrokes are clearly visible in the textures of the fields and the drifting steam, creating an atmosphere of tranquility that is both grand and full of nostalgic poetry.", + "seed": 14673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14673_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": -0.05268932103645056, + "vlm_logprob": -0.05268932103645056, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05268932103645056, + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12136444449424744 + ], + "opposite_logprob_tokens": [ + -2.2463643550872803 + ], + "logprob_reward": -0.12136444449424744, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12136444449424744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09573648124933243, + 0.07273055613040924, + 0.08054007589817047, + 0.043530214577913284, + 0.1570385843515396, + 0.14967025816440582, + 0.10396771878004074, + 0.16919644176959991, + 0.090247742831707, + 0.03734198212623596 + ], + "credit_map": [ + 0.013115828856825829, + 0.015620294958353043, + 0.012078882195055485, + 0.014060888439416885, + 0.01225702278316021, + 0.00837388914078474, + 0.008619090542197227, + 0.01161058247089386, + 0.015019905753433704, + 0.005525253247469664, + 0.007576538249850273, + 0.007102003786712885, + 0.0077360812574625015, + 0.0069327340461313725, + 0.004796748980879784, + 0.018041294068098068, + 0.01766250841319561, + 0.0050214482471346855, + 0.02617511712014675, + 0.008350668475031853, + 0.006455930415540934, + 0.0036815896164625883, + 0.00644640764221549, + 0.006746396422386169, + 0.005304240621626377, + 0.008185941725969315, + 0.007039673160761595, + 0.0055374871008098125, + 0.005404735915362835, + 0.004373381845653057, + 0.004143343307077885, + 0.0035414102021604776, + 0.03281239792704582, + 0.01249940600246191, + 0.051724452525377274, + 0.01566515490412712, + 0.00940708164125681, + 0.008138849399983883, + 0.01752007193863392, + 0.009271158836781979, + 0.01661025919020176, + 0.01792607083916664, + 0.021371319890022278, + 0.02157697267830372, + 0.015382800251245499, + 0.014675857499241829, + 0.018326440826058388, + 0.023800529539585114, + 0.01109041552990675, + 0.016965370625257492, + 0.01643347553908825, + 0.016191702336072922, + 0.010244368575513363, + 0.007558697368949652, + 0.014254330657422543, + 0.011229353956878185, + 0.01583557389676571, + 0.019263360649347305, + 0.030706288293004036, + 0.024697264656424522, + 0.01691805198788643, + 0.02070355787873268, + 0.016146261245012283, + 0.024926073849201202, + 0.010130493901669979, + 0.013357278890907764, + 0.01538985688239336, + 0.009913921356201172, + 0.010268471203744411, + 0.010639498010277748, + 0.00979149341583252, + 0.0107567198574543, + 0.006907809525728226, + 0.003151974640786648, + 0.005613663233816624, + 0.0050605288706719875, + 0.003974424209445715, + 0.00270716636441648, + 0.005129799246788025, + 0.004796614404767752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.12136444449424744 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294320970773697 + ], + "opposite_logprob_tokens": [ + -2.237943172454834 + ], + "logprob_reward": -0.11294320970773697, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294320970773697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1278572827577591, + 0.14888155460357666, + 0.12331617623567581, + 0.1362861543893814, + 0.13242869079113007, + 0.11404796689748764, + 0.04090389981865883, + 0.09884216636419296, + 0.026187563315033913, + 0.05124849081039429 + ], + "credit_map": [ + 0.021729901432991028, + 0.011364629492163658, + 0.014305481687188148, + 0.018092595040798187, + 0.015143188647925854, + 0.01716209016740322, + 0.01061834767460823, + 0.019441058859229088, + 0.03597404062747955, + 0.009866397827863693, + 0.011246377602219582, + 0.012723595835268497, + 0.011805806308984756, + 0.010327794589102268, + 0.0149201899766922, + 0.042017366737127304, + 0.02884676866233349, + 0.009294857271015644, + 0.029145868495106697, + 0.014413163997232914, + 0.014071229845285416, + 0.007464463356882334, + 0.009269794449210167, + 0.010810041800141335, + 0.021351240575313568, + 0.009601226076483727, + 0.03109058365225792, + 0.015658412128686905, + 0.024127155542373657, + 0.012971748597919941, + 0.011384475976228714, + 0.010101322084665298, + 0.02106691524386406, + 0.012092843651771545, + 0.03491692617535591, + 0.014717523008584976, + 0.013097688555717468, + 0.013046923093497753, + 0.01613975316286087, + 0.007350129075348377, + 0.012086613103747368, + 0.016006752848625183, + 0.01919456757605076, + 0.014344410970807076, + 0.013870935887098312, + 0.009110838174819946, + 0.016176683828234673, + 0.013257179409265518, + 0.00484099006280303, + 0.006584090180695057, + 0.0054296101443469524, + 0.004964328370988369, + 0.005087494384497404, + 0.004488283768296242, + 0.004998274613171816, + 0.004510832019150257, + 0.010315007530152798, + 0.014988567680120468, + 0.015374233946204185, + 0.017505308613181114, + 0.012440105900168419, + 0.00657802727073431, + 0.008848126046359539, + 0.0127927977591753, + 0.0026544483844190836, + 0.00309974723495543, + 0.00438599893823266, + 0.004618655890226364, + 0.003196149133145809, + 0.0019487547688186169, + 0.0029367406386882067, + 0.0033470718190073967, + 0.011813838966190815, + 0.004934967029839754, + 0.005517265293747187, + 0.0035121513064950705, + 0.0036248555406928062, + 0.0035708127543330193, + 0.004634202923625708, + 0.013640404678881168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.11294320970773697 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978199765086174 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.00978199765086174, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978199765086174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05437662452459335, + 0.053924012929201126, + 0.22024615108966827, + 0.04447411745786667, + 0.086006760597229, + 0.3451550006866455, + 0.032360538840293884, + 0.09100688248872757, + 0.027882901951670647, + 0.044567011296749115 + ], + "credit_map": [ + 0.00802932120859623, + 0.0064093065448105335, + 0.008206776343286037, + 0.008765404112637043, + 0.007950232364237309, + 0.0031981337815523148, + 0.005037624854594469, + 0.0067798239178955555, + 0.012947198003530502, + 0.005965552292764187, + 0.004050136543810368, + 0.0050560347735881805, + 0.004946501459926367, + 0.002551706274971366, + 0.003586224280297756, + 0.014820660464465618, + 0.044718410819768906, + 0.013473746366798878, + 0.06915595382452011, + 0.02041625790297985, + 0.024169284850358963, + 0.008492628112435341, + 0.018541431054472923, + 0.02127842977643013, + 0.00753045454621315, + 0.004862259142100811, + 0.01052157860249281, + 0.004659657366573811, + 0.005874109920114279, + 0.00364043191075325, + 0.004477022215723991, + 0.002908602124080062, + 0.01868835650384426, + 0.004191000945866108, + 0.02915049158036709, + 0.007524917833507061, + 0.004186497535556555, + 0.00653760926797986, + 0.01076013594865799, + 0.0049677579663693905, + 0.041950855404138565, + 0.02739662304520607, + 0.0522441528737545, + 0.042419422417879105, + 0.029061099514365196, + 0.05230795592069626, + 0.03364906460046768, + 0.06612581759691238, + 0.0036517793778330088, + 0.003368600970134139, + 0.0062841614708304405, + 0.004281739704310894, + 0.00365999573841691, + 0.0040456317365169525, + 0.003628803649917245, + 0.0034398273564875126, + 0.010208483785390854, + 0.00820914190262556, + 0.01771867275238037, + 0.009538174606859684, + 0.007368131075054407, + 0.012342346832156181, + 0.008936272002756596, + 0.01668565347790718, + 0.00313140288926661, + 0.004310664255172014, + 0.004347571637481451, + 0.0029747439548373222, + 0.0034940573386847973, + 0.002988505410030484, + 0.0033105304464697838, + 0.003325426485389471, + 0.007990389131009579, + 0.005629986058920622, + 0.005153842736035585, + 0.00462038628757, + 0.0038632501382380724, + 0.003859943011775613, + 0.0037193745374679565, + 0.009729837998747826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.00978199765086174 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011798333143815398 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011798333143815398, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011798333143815398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0952688455581665, + 0.08675991743803024, + 0.20050306618213654, + 0.15985658764839172, + 0.0674101710319519, + 0.059489838778972626, + 0.0700196772813797, + 0.08128217607736588, + 0.024376166984438896, + 0.15503357350826263 + ], + "credit_map": [ + 0.014836401678621769, + 0.013308124616742134, + 0.0076389131136238575, + 0.013903609476983547, + 0.012823511846363544, + 0.01301151979714632, + 0.006871241610497236, + 0.012875517830252647, + 0.022993452847003937, + 0.007121115457266569, + 0.006294058635830879, + 0.0070560043677687645, + 0.007571629248559475, + 0.0034316652454435825, + 0.006186565384268761, + 0.026105426251888275, + 0.045958876609802246, + 0.008701875805854797, + 0.06656494736671448, + 0.015449397265911102, + 0.013455228880047798, + 0.02408118173480034, + 0.01256060041487217, + 0.013730964623391628, + 0.040897589176893234, + 0.011569362133741379, + 0.05498525872826576, + 0.011376023292541504, + 0.008324261754751205, + 0.00805020984262228, + 0.013158682733774185, + 0.011495186947286129, + 0.014347449876368046, + 0.004007891286164522, + 0.020673274993896484, + 0.006463005207479, + 0.005683417897671461, + 0.00498760212212801, + 0.0073425122536718845, + 0.0039050148334354162, + 0.0069290692918002605, + 0.007123827002942562, + 0.008710415102541447, + 0.007955335080623627, + 0.006226147525012493, + 0.005492572672665119, + 0.007717329077422619, + 0.009335141628980637, + 0.007440242916345596, + 0.00731273926794529, + 0.01101208757609129, + 0.009606972336769104, + 0.0077278828248381615, + 0.010361568070948124, + 0.008747576735913754, + 0.007810596376657486, + 0.009637736715376377, + 0.011313035152852535, + 0.01257763709872961, + 0.011740602552890778, + 0.009525598026812077, + 0.005889230407774448, + 0.008343973197042942, + 0.012254350818693638, + 0.0027365584392100573, + 0.0030874975491315126, + 0.002920743077993393, + 0.003563069738447666, + 0.002849958138540387, + 0.0034840968437492847, + 0.002413491951301694, + 0.003320749383419752, + 0.03351905941963196, + 0.013420971110463142, + 0.017560012638568878, + 0.012603620998561382, + 0.00935002975165844, + 0.009041305631399155, + 0.01892613247036934, + 0.04061242938041687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.0011798333143815398 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817712001502514 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.01817712001502514, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817712001502514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25611400604248047, + 0.10563156753778458, + 0.048490654677152634, + 0.12896490097045898, + 0.08634108304977417, + 0.11652430146932602, + 0.07693928480148315, + 0.07672145217657089, + 0.04301820322871208, + 0.061254601925611496 + ], + "credit_map": [ + 0.031292613595724106, + 0.036531828343868256, + 0.035920947790145874, + 0.037724919617176056, + 0.037388529628515244, + 0.018831199035048485, + 0.026270730420947075, + 0.03215320408344269, + 0.015552405267953873, + 0.010536414571106434, + 0.010749473236501217, + 0.014003558084368706, + 0.014278905466198921, + 0.010938547551631927, + 0.0136133823543787, + 0.0159588735550642, + 0.009223861619830132, + 0.0032062139362096786, + 0.012626683339476585, + 0.006487166043370962, + 0.004361383151262999, + 0.003063586074858904, + 0.004305574577301741, + 0.005216184072196484, + 0.0332985520362854, + 0.005023141857236624, + 0.04716965928673744, + 0.010267418809235096, + 0.00828696507960558, + 0.006349172443151474, + 0.010609276592731476, + 0.007960716262459755, + 0.014050761237740517, + 0.00646108016371727, + 0.02403121069073677, + 0.011478723958134651, + 0.00876292772591114, + 0.005093140061944723, + 0.010386774316430092, + 0.006076458375900984, + 0.014117429964244366, + 0.014550501480698586, + 0.021421188488602638, + 0.010841223411262035, + 0.013960151933133602, + 0.010846210643649101, + 0.012599822133779526, + 0.01818777061998844, + 0.0065930490382015705, + 0.008663374930620193, + 0.016177356243133545, + 0.007819357328116894, + 0.011555952951312065, + 0.00734911160543561, + 0.011356408707797527, + 0.007424667477607727, + 0.007703944109380245, + 0.010280588641762733, + 0.01089107058942318, + 0.010746205225586891, + 0.008679088205099106, + 0.011516191065311432, + 0.00725526874884963, + 0.009649092331528664, + 0.004330012947320938, + 0.006344737019389868, + 0.007773672696202993, + 0.0052724662236869335, + 0.003890302497893572, + 0.005326703656464815, + 0.004329768940806389, + 0.005750535987317562, + 0.006573759485036135, + 0.006619961466640234, + 0.008792882785201073, + 0.011661672033369541, + 0.005796647630631924, + 0.006619465537369251, + 0.007002130150794983, + 0.008188078179955482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.01817712001502514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly panning drone shot captures the majestic sight of a vintage steam locomotive steadily and powerfully traversing a vast field. The entire scene takes on the quality of a dynamic oil painting, characterized by thick brushstrokes and vibrant colors. From above, the landscape resembles a vibrant canvas of golden wheat fields, emerald green meadows, and ochre soil, bisected by two parallel railroad tracks like a silver ribbon. A glossy black locomotive pulls several crimson and cream carriages along the straight track. Thick white steam billows from its chimney, leaving a long, gradually dissipating trail in the air, its form unmistakable, as if painted with a thick brush. The warm afternoon sun slants down, casting the train's long shadow across the field and reflecting a fluid sheen off its metal body. The colors are saturated and richly layered, and the brushstrokes are clearly visible in the textures of the fields and the drifting steam, creating an atmosphere of tranquility that is both grand and full of nostalgic poetry.", + "seed": 14674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_10/index_6387_seed_14674_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": -0.03503118231892586, + "vlm_logprob": -0.03503118231892586, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03503118231892586, + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07056884467601776 + ], + "opposite_logprob_tokens": [ + -2.820568799972534 + ], + "logprob_reward": -0.07056884467601776, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07056884467601776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08129507303237915, + 0.08672207593917847, + 0.02466236799955368, + 0.031310975551605225, + 0.17954884469509125, + 0.07644698768854141, + 0.054603081196546555, + 0.12035580724477768, + 0.0502505823969841, + 0.2948042154312134 + ], + "credit_map": [ + 0.011029678396880627, + 0.008396998047828674, + 0.014116537757217884, + 0.01956617459654808, + 0.007082503288984299, + 0.006729790009558201, + 0.0068032583221793175, + 0.0075701382011175156, + 0.010658183135092258, + 0.009269589558243752, + 0.015930162742733955, + 0.015232842415571213, + 0.006937009282410145, + 0.009198278188705444, + 0.012115960009396076, + 0.00738005293533206, + 0.0029311429243534803, + 0.0031319905538111925, + 0.004018926527351141, + 0.004160143900662661, + 0.0019021773478016257, + 0.0023737424053251743, + 0.004106488544493914, + 0.002037756610661745, + 0.0034809494391083717, + 0.003141164779663086, + 0.0034141961950808764, + 0.004909631330519915, + 0.004229960963129997, + 0.0060067372396588326, + 0.0031690907198935747, + 0.0029592455830425024, + 0.010558726266026497, + 0.009069454856216908, + 0.04625997319817543, + 0.07178206741809845, + 0.00911567360162735, + 0.009971324354410172, + 0.0102206626906991, + 0.012570973485708237, + 0.004715645685791969, + 0.006193576846271753, + 0.008175789378583431, + 0.01838885247707367, + 0.006837048567831516, + 0.010875646956264973, + 0.012930747121572495, + 0.008329682052135468, + 0.005334829445928335, + 0.006105819251388311, + 0.007558996323496103, + 0.006761603057384491, + 0.007369344122707844, + 0.00645177997648716, + 0.005503984168171883, + 0.00951672624796629, + 0.011503287591040134, + 0.012289540842175484, + 0.010335554368793964, + 0.00983862392604351, + 0.017339661717414856, + 0.028326012194156647, + 0.016220886260271072, + 0.014502239413559437, + 0.006235584616661072, + 0.007499285973608494, + 0.004641772247850895, + 0.007827153429389, + 0.005506532732397318, + 0.004692931659519672, + 0.007987191900610924, + 0.005860130302608013, + 0.013772149570286274, + 0.012267259880900383, + 0.01467958278954029, + 0.11355702579021454, + 0.009885040111839771, + 0.09082245081663132, + 0.021062148734927177, + 0.018758535385131836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.07056884467601776 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206229709088802 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206229709088802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206229709088802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11317717283964157, + 0.0437234528362751, + 0.03529547154903412, + 0.02554498240351677, + 0.034287843853235245, + 0.05492851510643959, + 0.07786403596401215, + 0.5666463375091553, + 0.020585864782333374, + 0.027946259826421738 + ], + "credit_map": [ + 0.008921211585402489, + 0.006856024265289307, + 0.030055468901991844, + 0.024617381393909454, + 0.00722638750448823, + 0.00883837416768074, + 0.00788436271250248, + 0.018777959048748016, + 0.006187026388943195, + 0.0036815644707530737, + 0.006090191658586264, + 0.005700537469238043, + 0.005670242477208376, + 0.0057253181003034115, + 0.005296294577419758, + 0.005372277926653624, + 0.0028867346700280905, + 0.00372805236838758, + 0.005981413647532463, + 0.006371856201440096, + 0.004685006570070982, + 0.00510905496776104, + 0.003308277577161789, + 0.0032250750809907913, + 0.0020724099595099688, + 0.0031045638024806976, + 0.0031074555590748787, + 0.004749781917780638, + 0.0030626202933490276, + 0.0036241328343749046, + 0.003440793603658676, + 0.00238322326913476, + 0.0029485104605555534, + 0.003643045900389552, + 0.00563559215515852, + 0.009466251358389854, + 0.0034955835435539484, + 0.0040908376686275005, + 0.0018869463820010424, + 0.0031210745219141245, + 0.003851386485621333, + 0.002374211559072137, + 0.0069666714407503605, + 0.0062517826445400715, + 0.002537341322749853, + 0.015611032955348492, + 0.002993466332554817, + 0.014342623762786388, + 0.005558484699577093, + 0.007185350172221661, + 0.012182757258415222, + 0.007791973650455475, + 0.017330674454569817, + 0.007156897336244583, + 0.00650592427700758, + 0.014151978306472301, + 0.031017757952213287, + 0.02575604058802128, + 0.028022296726703644, + 0.03073596954345703, + 0.0434926338493824, + 0.1860044151544571, + 0.1865236759185791, + 0.035093579441308975, + 0.0019238428212702274, + 0.003725769929587841, + 0.0014296500012278557, + 0.002382566686719656, + 0.0034381262958049774, + 0.003660244634374976, + 0.002219857182353735, + 0.0018058073474094272, + 0.0023275925777852535, + 0.00884107407182455, + 0.0014690071111544967, + 0.0019694985821843147, + 0.0023905739653855562, + 0.007653915788978338, + 0.0017322043422609568, + 0.001562394667416811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.014206229709088802 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187537789344788 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187537789344788, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187537789344788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22234652936458588, + 0.031807929277420044, + 0.023371582850813866, + 0.025318024680018425, + 0.035775426775217056, + 0.32887542247772217, + 0.019548706710338593, + 0.18196137249469757, + 0.013867119327187538, + 0.11712786555290222 + ], + "credit_map": [ + 0.01676999032497406, + 0.007971626706421375, + 0.08644313365221024, + 0.07446596026420593, + 0.008211825974285603, + 0.00921670999377966, + 0.011656536720693111, + 0.007610764354467392, + 0.004037467762827873, + 0.002804507501423359, + 0.0059804064221680164, + 0.005512705072760582, + 0.0023495727218687534, + 0.003699369030073285, + 0.004749740939587355, + 0.0026741614565253258, + 0.0023749328684061766, + 0.0025864236522465944, + 0.0046570864506065845, + 0.004506048280745745, + 0.0015150366816669703, + 0.0021468994673341513, + 0.003875355003401637, + 0.0017098025418817997, + 0.002364136977121234, + 0.0020655093248933554, + 0.003885614685714245, + 0.006010735407471657, + 0.0016002956544980407, + 0.002820252440869808, + 0.004504697397351265, + 0.0020667854696512222, + 0.0032371077686548233, + 0.0023951383773237467, + 0.0071910060942173, + 0.006373772397637367, + 0.003100879956036806, + 0.007323655765503645, + 0.0023096350487321615, + 0.003844229504466057, + 0.0268316138535738, + 0.014938609674572945, + 0.039884474128484726, + 0.033446524292230606, + 0.013451028615236282, + 0.09056217968463898, + 0.020164892077445984, + 0.08959610015153885, + 0.0023295353166759014, + 0.002775304252281785, + 0.002750363666564226, + 0.0015372935449704528, + 0.0034202132374048233, + 0.0020971850026398897, + 0.0019459344912320375, + 0.002692875685170293, + 0.009726948104798794, + 0.004825010895729065, + 0.0076431818306446075, + 0.0111539326608181, + 0.01033482700586319, + 0.06528589874505997, + 0.06173631548881531, + 0.01125527173280716, + 0.0015384978614747524, + 0.002608340932056308, + 0.0013076683972030878, + 0.001949057332240045, + 0.0016850029351189733, + 0.0026294414419680834, + 0.0010148525470867753, + 0.0011342575307935476, + 0.0067950026132166386, + 0.014690661802887917, + 0.005124981980770826, + 0.03355724364519119, + 0.004334553610533476, + 0.03974501043558121, + 0.006431650836020708, + 0.006448776461184025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.018187537789344788 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021977098658680916 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021977098658680916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021977098658680916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17769427597522736, + 0.05576842650771141, + 0.030781256034970284, + 0.02566884271800518, + 0.1514909565448761, + 0.15730689465999603, + 0.14627692103385925, + 0.12713252007961273, + 0.03970874473452568, + 0.08817116171121597 + ], + "credit_map": [ + 0.010946644470095634, + 0.007323572412133217, + 0.05785806477069855, + 0.0515773668885231, + 0.006690115202218294, + 0.0141111109405756, + 0.010092513635754585, + 0.01909489557147026, + 0.007082947064191103, + 0.005681999027729034, + 0.010917599312961102, + 0.009111094288527966, + 0.0066613866947591305, + 0.004070342052727938, + 0.007273890543729067, + 0.004969165660440922, + 0.0039774104952812195, + 0.004737745504826307, + 0.004551837220788002, + 0.0041809058748185635, + 0.0027917225379496813, + 0.0032105324789881706, + 0.004336179234087467, + 0.002994921989738941, + 0.0032441462390124798, + 0.0021073471289128065, + 0.002939952537417412, + 0.004743051249533892, + 0.0026565981097519398, + 0.004253612831234932, + 0.003249021712690592, + 0.0024751117452979088, + 0.008872823789715767, + 0.00664350064471364, + 0.04013286903500557, + 0.05892985686659813, + 0.007001888006925583, + 0.01124376617372036, + 0.008610986173152924, + 0.01005527563393116, + 0.010771069675683975, + 0.0061340960673987865, + 0.016596529632806778, + 0.00891414750367403, + 0.0063519589602947235, + 0.048051562160253525, + 0.014010248705744743, + 0.04647728055715561, + 0.01821199432015419, + 0.015560203231871128, + 0.02841937728226185, + 0.011291294358670712, + 0.01479271985590458, + 0.012643096968531609, + 0.012649521231651306, + 0.03270870819687843, + 0.010988712310791016, + 0.011361632496118546, + 0.011334906332194805, + 0.008898704312741756, + 0.017287207767367363, + 0.027836887165904045, + 0.030429668724536896, + 0.008994802832603455, + 0.0045533995144069195, + 0.007693316321820021, + 0.004224915057420731, + 0.005330544430762529, + 0.0039009246975183487, + 0.0053765298798680305, + 0.005113333463668823, + 0.0035157816018909216, + 0.011087744496762753, + 0.011188449338078499, + 0.01153475884348154, + 0.0162087120115757, + 0.009178047068417072, + 0.015144090168178082, + 0.007393519394099712, + 0.006435846444219351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0021977098658680916 + }, + { + "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 slow pan from a high aerial (drone) perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a glossy black vintage steam locomotive pulling crimson and cream carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a vast field with golden wheat, emerald green meadows, and ochre soil bisected by two parallel silver railroad tracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by thick brushstrokes and vibrant colors resembling a dynamic oil painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is thick white steam visible billowing from the locomotive's chimney and leaving a long, dissipating trail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999558955430984 + ], + "opposite_logprob_tokens": [ + -2.694995641708374 + ], + "logprob_reward": -0.06999558955430984, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999558955430984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20746281743049622, + 0.0740898922085762, + 0.023412100970745087, + 0.022343195974826813, + 0.07726424932479858, + 0.05363382026553154, + 0.13994824886322021, + 0.20654068887233734, + 0.06719916313886642, + 0.12810583412647247 + ], + "credit_map": [ + 0.01597590744495392, + 0.010184619575738907, + 0.06642314046621323, + 0.05376380309462547, + 0.012594497762620449, + 0.013526767492294312, + 0.011745225638151169, + 0.023248858749866486, + 0.008853832259774208, + 0.006918634753674269, + 0.013587388209998608, + 0.013096652925014496, + 0.006976430770009756, + 0.007745575159788132, + 0.010062860324978828, + 0.006848514080047607, + 0.001845476683229208, + 0.0026308009400963783, + 0.0032222929876297712, + 0.004682945087552071, + 0.003009248524904251, + 0.0031031123362481594, + 0.002583810593932867, + 0.002334414981305599, + 0.0018831525230780244, + 0.002205226570367813, + 0.003753117984160781, + 0.0032848429400473833, + 0.0024255397729575634, + 0.0035259348805993795, + 0.002869620453566313, + 0.0023957611992955208, + 0.004650115501135588, + 0.004189167637377977, + 0.016578948125243187, + 0.02957089990377426, + 0.005048949737101793, + 0.006248618941754103, + 0.005158316809684038, + 0.005819234065711498, + 0.004152808338403702, + 0.0031699533574283123, + 0.006646165158599615, + 0.012106702663004398, + 0.003961549606174231, + 0.0050522093661129475, + 0.013446122407913208, + 0.005098309833556414, + 0.009822847321629524, + 0.012417920865118504, + 0.033013444393873215, + 0.008789293467998505, + 0.019246164709329605, + 0.010018013417720795, + 0.01255830004811287, + 0.03408227488398552, + 0.012402764521539211, + 0.010372171178460121, + 0.010596911422908306, + 0.011029766872525215, + 0.016613459214568138, + 0.0690157487988472, + 0.06396450102329254, + 0.01254535373300314, + 0.008899728767573833, + 0.015804627910256386, + 0.006920618936419487, + 0.007951706647872925, + 0.008992777206003666, + 0.008724614977836609, + 0.005228386260569096, + 0.004676699638366699, + 0.012256404384970665, + 0.027537954971194267, + 0.008402158506214619, + 0.03892477974295616, + 0.00982182752341032, + 0.0099719213321805, + 0.011559522710740566, + 0.009631252847611904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.06999558955430984 + } + ], + "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/0000167/group_11/index_tensor([8739, 8739])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_tensor([8739, 8739])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a05629daafb1c840c3de3fb31d7a35440387f38 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_tensor([8739, 8739])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in the whimsical and vibrant style of '90s Disney animation, this static, eye-level medium close-up focuses on a grumpy, scholarly owl. The owl is plump, with meticulously detailed brown and white feathers, and a pair of small, round, gold-rimmed glasses perched precariously on its sharp, dark beak. As he hunches over a thick, dusty tome, his large, expressive golden eyes narrow in comic annoyance, his feathery brows furrowed in deep frustration. The book, bound in deep red leather with a richly textured, worn surface and faded gold foil filigree, rests on a dark wooden surface. With a swift, decisive, and distinctly impatient motion, the owl thrusts its beak downward, pecking fiercely at the book's cover. The impact sends a small, subtle puff of gray dust into the air, where the particles catch the warm, soft light illuminating the scene and swirl for a moment. The setting is a cozy, out-of-focus library, hinted at by the warm tones of the wooden bookshelves. The entire image is defined by crisp, clean hand-drawn lines and a saturated color palette, creating an atmosphere of charm, humor, and focus that perfectly captures the character's scholarly yet cantankerous personality.", + "seed": 3763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_8739_seed_3763_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": -0.008421980496495961, + "vlm_logprob": -0.008421980496495961, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.008421980496495961, + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003042832249775529 + ], + "opposite_logprob_tokens": [ + -7.503042697906494 + ], + "logprob_reward": -0.003042832249775529, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.003042832249775529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07486782222986221, + 0.029273677617311478, + 0.01294378936290741, + 0.02638312615454197, + 0.03805679455399513, + 0.011095554567873478, + 0.13061675429344177, + 0.6143893599510193, + 0.022034412249922752, + 0.04033873602747917 + ], + "credit_map": [ + 0.016498766839504242, + 0.004960835445672274, + 0.009532053954899311, + 0.007896693423390388, + 0.008580214343965054, + 0.00991376768797636, + 0.007877982221543789, + 0.009607512503862381, + 0.005529024172574282, + 0.0025834718253463507, + 0.0032751953694969416, + 0.003127503441646695, + 0.003897004993632436, + 0.0033949974458664656, + 0.0038592007476836443, + 0.0036072786897420883, + 0.002890904201194644, + 0.0010455395095050335, + 0.0019694408401846886, + 0.0014394245808944106, + 0.0013078262563794851, + 0.0016247204039245844, + 0.0012228293344378471, + 0.0014431049348786473, + 0.005817112047225237, + 0.0025023019406944513, + 0.0034829287324100733, + 0.0024506791960448027, + 0.003506520763039589, + 0.002768395934253931, + 0.0026932673063129187, + 0.0031619183719158173, + 0.007137557491660118, + 0.002598133170977235, + 0.008232464082539082, + 0.004552241414785385, + 0.003020200412720442, + 0.004251155536621809, + 0.0034927839878946543, + 0.004772260785102844, + 0.001749765477143228, + 0.0010155468480661511, + 0.001389291021041572, + 0.001890231273137033, + 0.0013626014115288854, + 0.0012268662685528398, + 0.0012177678290754557, + 0.0012434839736670256, + 0.024044275283813477, + 0.009902114048600197, + 0.02080998569726944, + 0.008900927379727364, + 0.013293519616127014, + 0.025576598942279816, + 0.012987521477043629, + 0.015101814642548561, + 0.05526280403137207, + 0.01276539545506239, + 0.11495532840490341, + 0.04269940406084061, + 0.030629532411694527, + 0.0658901259303093, + 0.25080791115760803, + 0.04137883707880974, + 0.002972551854327321, + 0.002507528057321906, + 0.0018332615727558732, + 0.0020777073223143816, + 0.0030465179588645697, + 0.003111161757260561, + 0.003969771787524223, + 0.0025159118231385946, + 0.005202567670494318, + 0.004847255069762468, + 0.0030914125964045525, + 0.005429273471236229, + 0.0073056528344750404, + 0.0046395412646234035, + 0.004960042890161276, + 0.004862990230321884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.003042832249775529 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002826031530275941 + ], + "opposite_logprob_tokens": [ + -5.87782621383667 + ], + "logprob_reward": -0.002826031530275941, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002826031530275941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163243368268013, + 0.14919553697109222, + 0.0718989297747612, + 0.08815208077430725, + 0.06499484181404114, + 0.08472484350204468, + 0.08266925811767578, + 0.1483093798160553, + 0.035053618252277374, + 0.11175812035799026 + ], + "credit_map": [ + 0.018758393824100494, + 0.016783449798822403, + 0.01244177296757698, + 0.01902051270008087, + 0.028894683346152306, + 0.01752198487520218, + 0.031366970390081406, + 0.018455620855093002, + 0.018925810232758522, + 0.01373294461518526, + 0.011103826574981213, + 0.018084684386849403, + 0.02164166234433651, + 0.017240354791283607, + 0.030326049774885178, + 0.018140215426683426, + 0.007654461544007063, + 0.0071639688685536385, + 0.004918655380606651, + 0.008964277803897858, + 0.011063458397984505, + 0.009130269289016724, + 0.013908413238823414, + 0.009095425717532635, + 0.014329035766422749, + 0.00799072440713644, + 0.013347071595489979, + 0.010772798210382462, + 0.012209577485918999, + 0.008204595185816288, + 0.009472946636378765, + 0.011825338006019592, + 0.010559894144535065, + 0.006566903088241816, + 0.006526915356516838, + 0.00895701628178358, + 0.008550191298127174, + 0.00611259788274765, + 0.008871984668076038, + 0.00884934701025486, + 0.0127396946772933, + 0.007702254690229893, + 0.010523352771997452, + 0.010280443355441093, + 0.016671273857355118, + 0.00806997437030077, + 0.009914739988744259, + 0.008823110722005367, + 0.009402078576385975, + 0.007326774764806032, + 0.006881524808704853, + 0.008523986674845219, + 0.016681550070643425, + 0.012774171307682991, + 0.007599124684929848, + 0.013480051420629025, + 0.01238690409809351, + 0.008266215212643147, + 0.01406841166317463, + 0.012275498360395432, + 0.01128434669226408, + 0.01967848278582096, + 0.05731486901640892, + 0.013034653849899769, + 0.002887725131586194, + 0.0035795969888567924, + 0.0025400700978934765, + 0.004001017659902573, + 0.006559022236615419, + 0.005329448729753494, + 0.0036326805129647255, + 0.00652405945584178, + 0.011460881680250168, + 0.016973866149783134, + 0.0068573602475225925, + 0.009148254990577698, + 0.027021374553442, + 0.01503797248005867, + 0.012064713053405285, + 0.013193701393902302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.002826031530275941 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00672908453270793 + ], + "opposite_logprob_tokens": [ + -5.0067291259765625 + ], + "logprob_reward": -0.00672908453270793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00672908453270793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05834190919995308, + 0.061670709401369095, + 0.03462769836187363, + 0.04321271926164627, + 0.10145989060401917, + 0.019606344401836395, + 0.30005982518196106, + 0.3240119218826294, + 0.01789756491780281, + 0.03911147266626358 + ], + "credit_map": [ + 0.010341580957174301, + 0.004181211348623037, + 0.007522820960730314, + 0.00702902814373374, + 0.006711011286824942, + 0.008633513003587723, + 0.0066855852492153645, + 0.0072371577844023705, + 0.008974500000476837, + 0.005335664842277765, + 0.006919838022440672, + 0.00829441286623478, + 0.010721538215875626, + 0.006880077999085188, + 0.00809953548014164, + 0.00644513638690114, + 0.005033921450376511, + 0.003182646119967103, + 0.003229998517781496, + 0.005162501707673073, + 0.0061539760790765285, + 0.0038240605499595404, + 0.004258714150637388, + 0.0037818769924342632, + 0.006879343651235104, + 0.004162134137004614, + 0.003963703289628029, + 0.005685839802026749, + 0.007784521672874689, + 0.004933444783091545, + 0.005229494534432888, + 0.004574238322675228, + 0.020537802949547768, + 0.007106834091246128, + 0.02181745506823063, + 0.009842248633503914, + 0.009215272963047028, + 0.010517284274101257, + 0.009954828768968582, + 0.012468156404793262, + 0.0034987146500498056, + 0.0015785149298608303, + 0.002586725866422057, + 0.002625845605507493, + 0.002407045103609562, + 0.002241624752059579, + 0.001877114875242114, + 0.0027907569892704487, + 0.055146362632513046, + 0.026385081931948662, + 0.0480567142367363, + 0.02061525732278824, + 0.0340055413544178, + 0.056693628430366516, + 0.02696072682738304, + 0.03219645842909813, + 0.02943260222673416, + 0.0232501532882452, + 0.026821302250027657, + 0.028825799003243446, + 0.03352174535393715, + 0.0528026781976223, + 0.09697171300649643, + 0.03238590061664581, + 0.002464116783812642, + 0.0014110793126747012, + 0.0013658312382176518, + 0.002706991508603096, + 0.0018090115627273917, + 0.001642612274736166, + 0.0038315339479595423, + 0.0026663884054869413, + 0.0048097581602633, + 0.004181297961622477, + 0.0035114805214107037, + 0.00621229549869895, + 0.005908838007599115, + 0.004838042426854372, + 0.0049026282504200935, + 0.004747126717120409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.00672908453270793 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026323901489377022 + ], + "opposite_logprob_tokens": [ + -3.6513237953186035 + ], + "logprob_reward": -0.026323901489377022, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026323901489377022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1029805913567543, + 0.07176854461431503, + 0.04076514393091202, + 0.05356062948703766, + 0.16344186663627625, + 0.04023168981075287, + 0.20472626388072968, + 0.14748325943946838, + 0.047158800065517426, + 0.127883180975914 + ], + "credit_map": [ + 0.023974480107426643, + 0.009444056078791618, + 0.01187177561223507, + 0.013112499378621578, + 0.013781584799289703, + 0.012481903657317162, + 0.007027331739664078, + 0.011286971159279346, + 0.012524379417300224, + 0.004931210540235043, + 0.007638412527740002, + 0.009509571827948093, + 0.009734087623655796, + 0.010219714604318142, + 0.0081139812245965, + 0.009097187779843807, + 0.008128582499921322, + 0.0029739609453827143, + 0.005765234120190144, + 0.0049359058029949665, + 0.005329437088221312, + 0.004668715409934521, + 0.003481503576040268, + 0.005481807515025139, + 0.010778079740703106, + 0.004999115597456694, + 0.006316053215414286, + 0.006640204694122076, + 0.005536957643926144, + 0.0069963810965418816, + 0.0050444756634533405, + 0.0072493646293878555, + 0.0367763377726078, + 0.01083588320761919, + 0.03583893924951553, + 0.015552396886050701, + 0.01263966504484415, + 0.017057759687304497, + 0.01338384859263897, + 0.02135705016553402, + 0.008433012291789055, + 0.0030725831165909767, + 0.0043412852101027966, + 0.0045920987613499165, + 0.00502240564674139, + 0.004577776417136192, + 0.004950934089720249, + 0.005241598468273878, + 0.02635142393410206, + 0.018567729741334915, + 0.026958219707012177, + 0.021141067147254944, + 0.034841664135456085, + 0.02574813738465309, + 0.03038913570344448, + 0.020728901028633118, + 0.013937581330537796, + 0.012464907951653004, + 0.010358138009905815, + 0.012221965938806534, + 0.013658839277923107, + 0.025120917707681656, + 0.046444233506917953, + 0.013276690617203712, + 0.005864370614290237, + 0.005862024612724781, + 0.008685562759637833, + 0.0049274819903075695, + 0.006400054320693016, + 0.006229777354747057, + 0.004501014947891235, + 0.004688517656177282, + 0.01449174527078867, + 0.009483092464506626, + 0.009852019138634205, + 0.02684740349650383, + 0.014034155756235123, + 0.01928020641207695, + 0.015148443169891834, + 0.018746132031083107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.026323901489377022 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031880526803433895 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.0031880526803433895, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031880526803433895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05873855575919151, + 0.030481908470392227, + 0.012074700556695461, + 0.035932887345552444, + 0.05422654747962952, + 0.01548690628260374, + 0.1418289840221405, + 0.5880951881408691, + 0.021062595769762993, + 0.042071741074323654 + ], + "credit_map": [ + 0.011193937622010708, + 0.005171242170035839, + 0.007930085994303226, + 0.006112429779022932, + 0.007290167734026909, + 0.006915577221661806, + 0.006324900779873133, + 0.007800211198627949, + 0.005949181970208883, + 0.002381101716309786, + 0.00258921948261559, + 0.004152471199631691, + 0.0036299056373536587, + 0.004325097892433405, + 0.003914665896445513, + 0.0035402614157646894, + 0.0021233020815998316, + 0.0010221018455922604, + 0.0014493769267573953, + 0.001679210108704865, + 0.0014144729357212782, + 0.0014650820521637797, + 0.0013420311734080315, + 0.0015791227342560887, + 0.008391729556024075, + 0.0023581674322485924, + 0.006857686210423708, + 0.004111341200768948, + 0.0023533154744654894, + 0.004264921881258488, + 0.00248697423376143, + 0.0051087490282952785, + 0.01172569114714861, + 0.003801675047725439, + 0.011736766435205936, + 0.006401072721928358, + 0.0032728612422943115, + 0.005914388690143824, + 0.0036186256911605597, + 0.007755462545901537, + 0.002304073888808489, + 0.0017239408334717155, + 0.0016525447135791183, + 0.0016221324913203716, + 0.0030874607618898153, + 0.0015791887417435646, + 0.0016734423115849495, + 0.0018441216088831425, + 0.027141856029629707, + 0.01304401084780693, + 0.027915820479393005, + 0.011649440973997116, + 0.012198271229863167, + 0.017144445329904556, + 0.010198990814387798, + 0.0225361380726099, + 0.05451064929366112, + 0.012831591069698334, + 0.10968529433012009, + 0.04165330156683922, + 0.029923049733042717, + 0.06229471415281296, + 0.2386847287416458, + 0.03851182386279106, + 0.0020909684244543314, + 0.002287439536303282, + 0.0024640525225549936, + 0.002099777804687619, + 0.003052345011383295, + 0.002863967791199684, + 0.002992811845615506, + 0.0032112328335642815, + 0.004740684758871794, + 0.005069674924015999, + 0.003017149167135358, + 0.004807444289326668, + 0.008564039133489132, + 0.005438971798866987, + 0.005233247298747301, + 0.005200525280088186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0031880526803433895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in the whimsical and vibrant style of '90s Disney animation, this static, eye-level medium close-up focuses on a grumpy, scholarly owl. The owl is plump, with meticulously detailed brown and white feathers, and a pair of small, round, gold-rimmed glasses perched precariously on its sharp, dark beak. As he hunches over a thick, dusty tome, his large, expressive golden eyes narrow in comic annoyance, his feathery brows furrowed in deep frustration. The book, bound in deep red leather with a richly textured, worn surface and faded gold foil filigree, rests on a dark wooden surface. With a swift, decisive, and distinctly impatient motion, the owl thrusts its beak downward, pecking fiercely at the book's cover. The impact sends a small, subtle puff of gray dust into the air, where the particles catch the warm, soft light illuminating the scene and swirl for a moment. The setting is a cozy, out-of-focus library, hinted at by the warm tones of the wooden bookshelves. The entire image is defined by crisp, clean hand-drawn lines and a saturated color palette, creating an atmosphere of charm, humor, and focus that perfectly captures the character's scholarly yet cantankerous personality.", + "seed": 3764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_8739_seed_3764_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": -0.011198841035366058, + "vlm_logprob": -0.011198841035366058, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.011198841035366058, + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00204889546148479 + ], + "opposite_logprob_tokens": [ + -7.252048969268799 + ], + "logprob_reward": -0.00204889546148479, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.00204889546148479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13000711798667908, + 0.12414439022541046, + 0.07685578614473343, + 0.06080559268593788, + 0.029769491404294968, + 0.04535379633307457, + 0.07992206513881683, + 0.1372254192829132, + 0.09114404022693634, + 0.22477228939533234 + ], + "credit_map": [ + 0.01770281046628952, + 0.006564106326550245, + 0.006232377607375383, + 0.014834431000053883, + 0.010778463445603848, + 0.012930815108120441, + 0.04654601961374283, + 0.014418094418942928, + 0.021372592076659203, + 0.0050236922688782215, + 0.0038367200177162886, + 0.016033563762903214, + 0.006002587266266346, + 0.012983555905520916, + 0.05115329846739769, + 0.007738396525382996, + 0.006785952020436525, + 0.0037853603716939688, + 0.0025598141364753246, + 0.0058153895661234856, + 0.008561193011701107, + 0.02105885185301304, + 0.024383407086133957, + 0.0039058197289705276, + 0.011943094432353973, + 0.006741897203028202, + 0.007581030484288931, + 0.006309932097792625, + 0.007317584939301014, + 0.007508685812354088, + 0.00550704263150692, + 0.00789632648229599, + 0.004089637193828821, + 0.003399243811145425, + 0.0021008651237934828, + 0.002524401992559433, + 0.005757416132837534, + 0.0035740791354328394, + 0.00323208374902606, + 0.005091765429824591, + 0.007917765527963638, + 0.004175882320851088, + 0.005958732683211565, + 0.00934625044465065, + 0.004343072883784771, + 0.003745565889403224, + 0.004096810705959797, + 0.005769719835370779, + 0.009560314938426018, + 0.014038007706403732, + 0.00519557436928153, + 0.007290365174412727, + 0.017711741849780083, + 0.006531294900923967, + 0.011686709709465504, + 0.007908059284090996, + 0.017550362274050713, + 0.029291072860360146, + 0.012561938725411892, + 0.02274380251765251, + 0.012789478525519371, + 0.017322996631264687, + 0.01089478563517332, + 0.014070986770093441, + 0.009314123541116714, + 0.032068412750959396, + 0.003070438513532281, + 0.005522247403860092, + 0.005610637832432985, + 0.023850716650485992, + 0.004240269772708416, + 0.007467200513929129, + 0.030837925150990486, + 0.053796183317899704, + 0.010109263472259045, + 0.015412973240017891, + 0.01916392333805561, + 0.0684446170926094, + 0.011549676768481731, + 0.01545773446559906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.00204889546148479 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640998784452677 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640998784452677, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640998784452677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09092232584953308, + 0.03204147517681122, + 0.08891107887029648, + 0.3187922239303589, + 0.054266758263111115, + 0.1315595954656601, + 0.0886835902929306, + 0.03708680719137192, + 0.09766896069049835, + 0.060067158192396164 + ], + "credit_map": [ + 0.016957590356469154, + 0.008990276604890823, + 0.010356073267757893, + 0.012711809948086739, + 0.009447015821933746, + 0.008820798248052597, + 0.00965941697359085, + 0.013979343697428703, + 0.00640011066570878, + 0.0027800092939287424, + 0.0022666864097118378, + 0.0035734977573156357, + 0.0030948775820434093, + 0.004993559792637825, + 0.004997465759515762, + 0.003935270942747593, + 0.007785334717482328, + 0.004803178366273642, + 0.0029109346214681864, + 0.007279407698661089, + 0.013055176474153996, + 0.020241912454366684, + 0.02784944884479046, + 0.00498569430783391, + 0.0866115465760231, + 0.018659498542547226, + 0.020736373960971832, + 0.01655297912657261, + 0.026456689462065697, + 0.11450187861919403, + 0.012296704575419426, + 0.022976551204919815, + 0.015067261643707752, + 0.0035807781387120485, + 0.0028415166307240725, + 0.003907348029315472, + 0.003534063696861267, + 0.004222570452839136, + 0.0037302973214536905, + 0.017382925376296043, + 0.01782681979238987, + 0.009363498538732529, + 0.01632046326994896, + 0.024624578654766083, + 0.013618837110698223, + 0.014946901239454746, + 0.014511252753436565, + 0.020347241312265396, + 0.009904514998197556, + 0.017299842089414597, + 0.0069406526163220406, + 0.007560296915471554, + 0.016304954886436462, + 0.007017842028290033, + 0.014008736237883568, + 0.009646749123930931, + 0.004418558906763792, + 0.0054451534524559975, + 0.0026805177330970764, + 0.0034542405046522617, + 0.006233806721866131, + 0.0053686900064349174, + 0.004695820156484842, + 0.0047900197096168995, + 0.010754377581179142, + 0.030252279713749886, + 0.0038089470472186804, + 0.005455667153000832, + 0.006372764240950346, + 0.02839263714849949, + 0.0063686068169772625, + 0.006263676565140486, + 0.00873349979519844, + 0.007923676632344723, + 0.005165623035281897, + 0.005993202794343233, + 0.005637818481773138, + 0.014720331877470016, + 0.007832014933228493, + 0.004060991108417511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.004640998784452677 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621141150593758 + ], + "opposite_logprob_tokens": [ + -4.8826212882995605 + ], + "logprob_reward": -0.007621141150593758, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621141150593758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125608891248703, + 0.04404854029417038, + 0.05966296046972275, + 0.16010130941867828, + 0.04319538176059723, + 0.040964387357234955, + 0.05820145085453987, + 0.03394439443945885, + 0.3908126950263977, + 0.043459996581077576 + ], + "credit_map": [ + 0.02336171455681324, + 0.007529580499976873, + 0.006602809298783541, + 0.013481049798429012, + 0.00943240150809288, + 0.011040464974939823, + 0.04210589453577995, + 0.012054980732500553, + 0.005346618592739105, + 0.0017963004065677524, + 0.0013444562209770083, + 0.006311391014605761, + 0.002143294084817171, + 0.007315656635910273, + 0.017288440838456154, + 0.002502382965758443, + 0.004524989519268274, + 0.0019037487218156457, + 0.0016986782429739833, + 0.004997046198695898, + 0.003621114417910576, + 0.01815102808177471, + 0.021564621478319168, + 0.003201734274625778, + 0.04056362807750702, + 0.009785822592675686, + 0.009563511237502098, + 0.006190828047692776, + 0.015662791207432747, + 0.05807892233133316, + 0.007226122543215752, + 0.013029674999415874, + 0.006817707326263189, + 0.0027976541314274073, + 0.0026778224855661392, + 0.0030005252920091152, + 0.00510128540918231, + 0.0029746834188699722, + 0.007523919455707073, + 0.012301784008741379, + 0.0061433748342096806, + 0.0027084415778517723, + 0.007870867848396301, + 0.009380673989653587, + 0.003468555398285389, + 0.002825254574418068, + 0.0030004309955984354, + 0.005566786974668503, + 0.005452385172247887, + 0.007439854089170694, + 0.005587511695921421, + 0.006089746952056885, + 0.01343563199043274, + 0.005488574970513582, + 0.008869914337992668, + 0.005837828386574984, + 0.003893817774951458, + 0.0050116064958274364, + 0.003344072261825204, + 0.005945115815848112, + 0.003516498254612088, + 0.003882796037942171, + 0.004031939432024956, + 0.004318551160395145, + 0.03858160972595215, + 0.13466176390647888, + 0.012693529017269611, + 0.01897524856030941, + 0.026684587821364403, + 0.11253482103347778, + 0.02056220732629299, + 0.026118919253349304, + 0.004349226132035255, + 0.00797353032976389, + 0.003408884396776557, + 0.00451515382155776, + 0.005367385223507881, + 0.009586363099515438, + 0.00535482307896018, + 0.0029046337585896254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.007621141150593758 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807205334305763 + ], + "opposite_logprob_tokens": [ + -3.288072109222412 + ], + "logprob_reward": -0.03807205334305763, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807205334305763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1429174542427063, + 0.13438281416893005, + 0.1452363133430481, + 0.11878875643014908, + 0.04923931881785393, + 0.049117960035800934, + 0.087188720703125, + 0.057206891477108, + 0.09075848013162613, + 0.125163272023201 + ], + "credit_map": [ + 0.022469939664006233, + 0.006421197671443224, + 0.007398291490972042, + 0.014407958835363388, + 0.009258903563022614, + 0.017763961106538773, + 0.04889324679970741, + 0.016303962096571922, + 0.019611049443483353, + 0.005372688639909029, + 0.004588941112160683, + 0.01914096437394619, + 0.005790827330201864, + 0.018054677173495293, + 0.05419114604592323, + 0.007632523309439421, + 0.011695663444697857, + 0.0046042813919484615, + 0.004134221933782101, + 0.011971990577876568, + 0.007594414986670017, + 0.043691132217645645, + 0.05496453493833542, + 0.006580085027962923, + 0.029336361214518547, + 0.007997365668416023, + 0.0077243163250386715, + 0.008617271669209003, + 0.014631742611527443, + 0.03321997448801994, + 0.008838447742164135, + 0.008423284627497196, + 0.010590532794594765, + 0.002933437004685402, + 0.0033077532425522804, + 0.0030401491094380617, + 0.004704935476183891, + 0.0033156953286379576, + 0.005980440881103277, + 0.01536637358367443, + 0.006490308791399002, + 0.0030591648537665606, + 0.006671462673693895, + 0.013823945075273514, + 0.003906197613105178, + 0.003231578040868044, + 0.003618417540565133, + 0.008316883817315102, + 0.008032726123929024, + 0.011484132148325443, + 0.00623646704480052, + 0.008909988217055798, + 0.021029390394687653, + 0.008862190879881382, + 0.013451467268168926, + 0.009182355366647243, + 0.007000095676630735, + 0.008948476985096931, + 0.005133410915732384, + 0.009493758901953697, + 0.00616734242066741, + 0.007800712715834379, + 0.004810728598386049, + 0.007852362468838692, + 0.008178627118468285, + 0.027658142149448395, + 0.0036567896604537964, + 0.006280472502112389, + 0.009608152322471142, + 0.02259276993572712, + 0.0061032273806631565, + 0.00668030371889472, + 0.011384685523808002, + 0.03567364066839218, + 0.006483187898993492, + 0.008888643234968185, + 0.012109730392694473, + 0.03453553840517998, + 0.009135942906141281, + 0.0069518969394266605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03807205334305763 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611116437241435 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.003611116437241435, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611116437241435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20825417339801788, + 0.08001205325126648, + 0.025619911029934883, + 0.13204102218151093, + 0.02967079170048237, + 0.2159876823425293, + 0.04337252676486969, + 0.07055061310529709, + 0.13941390812397003, + 0.05507737398147583 + ], + "credit_map": [ + 0.04036519303917885, + 0.009281782433390617, + 0.009940257295966148, + 0.021725302562117577, + 0.010590008459985256, + 0.024096893146634102, + 0.07344266027212143, + 0.018812060356140137, + 0.013455895707011223, + 0.0034782150760293007, + 0.002974751405417919, + 0.012452791444957256, + 0.00429963693022728, + 0.006397351156920195, + 0.02992725372314453, + 0.007026155013591051, + 0.004322192631661892, + 0.0018037009285762906, + 0.0014044332783669233, + 0.003078895155340433, + 0.002860767999663949, + 0.0038147899322211742, + 0.006347071845084429, + 0.0019880591426044703, + 0.024986507371068, + 0.007768598385155201, + 0.008451681584119797, + 0.014047665521502495, + 0.011915437877178192, + 0.04189838096499443, + 0.006964510306715965, + 0.016008218750357628, + 0.004264097195118666, + 0.0021076819393783808, + 0.0023231753148138523, + 0.003486393252387643, + 0.0028320769779384136, + 0.00398043030872941, + 0.004501692019402981, + 0.006175242830067873, + 0.056357886642217636, + 0.012718181125819683, + 0.026517847552895546, + 0.050355516374111176, + 0.00871932040899992, + 0.014258434996008873, + 0.009645981714129448, + 0.03741449490189552, + 0.004144661594182253, + 0.006062070373445749, + 0.0031005756463855505, + 0.00436021201312542, + 0.010302738286554813, + 0.004045148380100727, + 0.006300003733485937, + 0.005057118367403746, + 0.008791192434728146, + 0.00999797135591507, + 0.006514310371130705, + 0.01025143638253212, + 0.008446984924376011, + 0.009668685495853424, + 0.007070683874189854, + 0.00980934128165245, + 0.014084049500524998, + 0.048646215349435806, + 0.005412613973021507, + 0.013511363416910172, + 0.009056142531335354, + 0.0326642282307148, + 0.006694393698126078, + 0.009344898164272308, + 0.006025233771651983, + 0.015957579016685486, + 0.0026790613774210215, + 0.0048878248780965805, + 0.007971059530973434, + 0.008826516568660736, + 0.005480428691953421, + 0.0032496617641299963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.003611116437241435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in the whimsical and vibrant style of '90s Disney animation, this static, eye-level medium close-up focuses on a grumpy, scholarly owl. The owl is plump, with meticulously detailed brown and white feathers, and a pair of small, round, gold-rimmed glasses perched precariously on its sharp, dark beak. As he hunches over a thick, dusty tome, his large, expressive golden eyes narrow in comic annoyance, his feathery brows furrowed in deep frustration. The book, bound in deep red leather with a richly textured, worn surface and faded gold foil filigree, rests on a dark wooden surface. With a swift, decisive, and distinctly impatient motion, the owl thrusts its beak downward, pecking fiercely at the book's cover. The impact sends a small, subtle puff of gray dust into the air, where the particles catch the warm, soft light illuminating the scene and swirl for a moment. The setting is a cozy, out-of-focus library, hinted at by the warm tones of the wooden bookshelves. The entire image is defined by crisp, clean hand-drawn lines and a saturated color palette, creating an atmosphere of charm, humor, and focus that perfectly captures the character's scholarly yet cantankerous personality.", + "seed": 3765, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_8739_seed_3765_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": -0.014642621111124753, + "vlm_logprob": -0.014642621111124753, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.014642621111124753, + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006982567720115185 + ], + "opposite_logprob_tokens": [ + -6.881982803344727 + ], + "logprob_reward": -0.006982567720115185, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.006982567720115185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07501666247844696, + 0.04255056753754616, + 0.019234364852309227, + 0.01741529069840908, + 0.6268260478973389, + 0.0333816334605217, + 0.06849026679992676, + 0.034606385976076126, + 0.046404607594013214, + 0.036074165254831314 + ], + "credit_map": [ + 0.01356640737503767, + 0.005996981635689735, + 0.005001494195312262, + 0.004029123578220606, + 0.004440427292138338, + 0.012987778522074223, + 0.021412499248981476, + 0.007581952027976513, + 0.00958013255149126, + 0.0036935554817318916, + 0.0026859247591346502, + 0.0028338106349110603, + 0.003958301153033972, + 0.002803516574203968, + 0.013907750137150288, + 0.0030875802040100098, + 0.0019671027548611164, + 0.002232250990346074, + 0.0019275897648185492, + 0.0018193116411566734, + 0.004152464680373669, + 0.0029446796979755163, + 0.002405187115073204, + 0.0017857782077044249, + 0.0024754281621426344, + 0.001981077017262578, + 0.0013614385388791561, + 0.0010609050514176488, + 0.002577941631898284, + 0.0024323889520019293, + 0.0028706740122288465, + 0.0026554365176707506, + 0.1853598952293396, + 0.052503496408462524, + 0.03347478806972504, + 0.22988249361515045, + 0.025068169459700584, + 0.023324383422732353, + 0.034636422991752625, + 0.04257640242576599, + 0.0037889620289206505, + 0.003703230293467641, + 0.0027822251431643963, + 0.003570364788174629, + 0.004987447056919336, + 0.005380641203373671, + 0.0055432626977562904, + 0.0036255016457289457, + 0.011045824736356735, + 0.008995207026600838, + 0.0057431841269135475, + 0.017722396180033684, + 0.008677155710756779, + 0.006516141351312399, + 0.005487545859068632, + 0.0043028127402067184, + 0.0030083272140473127, + 0.0069790189154446125, + 0.002484329743310809, + 0.0030024172738194466, + 0.005215139593929052, + 0.00648118369281292, + 0.004406007006764412, + 0.0030299611389636993, + 0.006474349647760391, + 0.009393485262989998, + 0.0027995542623102665, + 0.004788331687450409, + 0.009549803100526333, + 0.0045069484040141106, + 0.005622923374176025, + 0.0032692088279873133, + 0.004097347613424063, + 0.004964545834809542, + 0.003877356881275773, + 0.00343904085457325, + 0.003936916124075651, + 0.008809518069028854, + 0.003233806462958455, + 0.0037156324833631516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.006982567720115185 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00221055606380105 + ], + "opposite_logprob_tokens": [ + -6.12721061706543 + ], + "logprob_reward": -0.00221055606380105, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00221055606380105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04984041675925255, + 0.03825693577528, + 0.018144749104976654, + 0.017170444130897522, + 0.5902760624885559, + 0.06497801095247269, + 0.12151695787906647, + 0.016202412545681, + 0.04184185341000557, + 0.04177211970090866 + ], + "credit_map": [ + 0.009607204236090183, + 0.004902759101241827, + 0.003079572692513466, + 0.0030517461709678173, + 0.004309612791985273, + 0.007068458944559097, + 0.013941890560090542, + 0.00387917784973979, + 0.004176784306764603, + 0.0026267811190336943, + 0.0024352155160158873, + 0.003446694929152727, + 0.005125745199620724, + 0.004346338100731373, + 0.012896916829049587, + 0.0032024644315242767, + 0.00216692965477705, + 0.0016795824049040675, + 0.0019820870365947485, + 0.0014110883930698037, + 0.002602190477773547, + 0.0038898454513400793, + 0.002602157648652792, + 0.0018108714139088988, + 0.002076422329992056, + 0.002755742287263274, + 0.0011876750504598022, + 0.0013590988237410784, + 0.00284180068410933, + 0.0025947766844183207, + 0.00266848411411047, + 0.0016864463686943054, + 0.17598967254161835, + 0.04875722900032997, + 0.031553566455841064, + 0.21895147860050201, + 0.02179248072206974, + 0.018832217901945114, + 0.03327970579266548, + 0.0411197803914547, + 0.009238891303539276, + 0.005298910662531853, + 0.004858624655753374, + 0.006820420268923044, + 0.011232898570597172, + 0.009938877075910568, + 0.012077647261321545, + 0.005511743947863579, + 0.022597787901759148, + 0.01190478727221489, + 0.008319026790559292, + 0.03788295388221741, + 0.012377412989735603, + 0.009784433990716934, + 0.010677080601453781, + 0.007973487488925457, + 0.001363903982564807, + 0.002910377224907279, + 0.0010420323815196753, + 0.0017183617455884814, + 0.002783440751954913, + 0.0035565446596592665, + 0.0015799703542143106, + 0.0012477838899940252, + 0.004940610844641924, + 0.004504684824496508, + 0.004146846942603588, + 0.003885995363816619, + 0.007263233885169029, + 0.006387763191014528, + 0.006241363938897848, + 0.004471360705792904, + 0.005338564049452543, + 0.007357221562415361, + 0.003384601091966033, + 0.004227790515869856, + 0.007584302220493555, + 0.006015062797814608, + 0.00466378778219223, + 0.003200796665623784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.00221055606380105 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056759223341942 + ], + "opposite_logprob_tokens": [ + -3.8955676555633545 + ], + "logprob_reward": -0.02056759223341942, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056759223341942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0716978907585144, + 0.07623662799596786, + 0.0640392005443573, + 0.021007422357797623, + 0.616705596446991, + 0.030818352475762367, + 0.057299766689538956, + 0.01462920755147934, + 0.024835148826241493, + 0.022730832919478416 + ], + "credit_map": [ + 0.01567336916923523, + 0.0041663129813969135, + 0.004029391333460808, + 0.004203418735414743, + 0.0031524179503321648, + 0.004782299976795912, + 0.029760049656033516, + 0.005930629558861256, + 0.016544653102755547, + 0.003945573698729277, + 0.0038312713149935007, + 0.004032928962260485, + 0.003804584266617894, + 0.005106453318148851, + 0.032443951815366745, + 0.006527202669531107, + 0.014628209173679352, + 0.002765790093690157, + 0.004540683701634407, + 0.003301351796835661, + 0.002934078685939312, + 0.004514936823397875, + 0.025608204305171967, + 0.005745946429669857, + 0.0018172742566093802, + 0.0028600289952009916, + 0.0014483120758086443, + 0.0012676175683736801, + 0.004948384128510952, + 0.0026403069496154785, + 0.003990474157035351, + 0.0020350227132439613, + 0.18309971690177917, + 0.051361486315727234, + 0.0327971875667572, + 0.22912339866161346, + 0.02254907414317131, + 0.020061034709215164, + 0.034786056727170944, + 0.04292765632271767, + 0.004263902548700571, + 0.0029111530166119337, + 0.002024404238909483, + 0.002994084730744362, + 0.005448135081678629, + 0.004938021767884493, + 0.0055825477465987206, + 0.002656102878972888, + 0.007404966279864311, + 0.0060136024840176105, + 0.0029288241639733315, + 0.010692798532545567, + 0.012540129944682121, + 0.007167303003370762, + 0.007318986114114523, + 0.0032331570982933044, + 0.0013334453105926514, + 0.002965937601402402, + 0.0011378119233995676, + 0.001141602755524218, + 0.001962411915883422, + 0.003034133929759264, + 0.001922322902828455, + 0.0011315419105812907, + 0.004061955958604813, + 0.004602715373039246, + 0.002660187892615795, + 0.0022069597616791725, + 0.001989261945709586, + 0.0031365584582090378, + 0.0029815565794706345, + 0.0031959530897438526, + 0.0028076847083866596, + 0.004472340457141399, + 0.0021785811986774206, + 0.0022013511043041945, + 0.0032997603993862867, + 0.002704781712964177, + 0.002733325120061636, + 0.0023330082185566425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.02056759223341942 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367723897099495 + ], + "opposite_logprob_tokens": [ + -3.408677339553833 + ], + "logprob_reward": -0.03367723897099495, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367723897099495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08313973993062973, + 0.04916685074567795, + 0.0257140900939703, + 0.058757416903972626, + 0.4004781246185303, + 0.09189031273126602, + 0.13816827535629272, + 0.020532455295324326, + 0.08359618484973907, + 0.04855657368898392 + ], + "credit_map": [ + 0.012460112571716309, + 0.006290987599641085, + 0.005740614142268896, + 0.006822120398283005, + 0.008025571703910828, + 0.008500918745994568, + 0.024616364389657974, + 0.010683049447834492, + 0.010091547854244709, + 0.004529927857220173, + 0.00348954601213336, + 0.0035282985772937536, + 0.0048223198391497135, + 0.005320047028362751, + 0.011376239359378815, + 0.006008924450725317, + 0.005186923313885927, + 0.0023687737993896008, + 0.0031907125376164913, + 0.0019164172699674964, + 0.002021499676629901, + 0.004266052972525358, + 0.003134858561679721, + 0.0036288509145379066, + 0.005606281571090221, + 0.006194133777171373, + 0.003956647124141455, + 0.003753027180209756, + 0.012419437989592552, + 0.00950400810688734, + 0.009397606365382671, + 0.00792627315968275, + 0.11666392534971237, + 0.030531957745552063, + 0.02084851823747158, + 0.14904119074344635, + 0.017394043505191803, + 0.0133652463555336, + 0.02411610446870327, + 0.028517158702015877, + 0.01332869939506054, + 0.008133402094244957, + 0.0062443106435239315, + 0.008823360316455364, + 0.017223140224814415, + 0.013925179839134216, + 0.016722673550248146, + 0.007489547599107027, + 0.018980098888278008, + 0.012427865527570248, + 0.008489727973937988, + 0.03030439466238022, + 0.02905546873807907, + 0.013236640021204948, + 0.017104623839259148, + 0.00856947060674429, + 0.0017140923300758004, + 0.0026869410648941994, + 0.0017004882683977485, + 0.0023429172579199076, + 0.00394005561247468, + 0.0023207024205476046, + 0.0036668486427515745, + 0.0021604085341095924, + 0.011101508513092995, + 0.026810236275196075, + 0.007550742477178574, + 0.005689182784408331, + 0.006695752032101154, + 0.013267949223518372, + 0.00674430700019002, + 0.005736507475376129, + 0.006590454373508692, + 0.005103397183120251, + 0.0030481989961117506, + 0.006001413334161043, + 0.0031979165505617857, + 0.004374187905341387, + 0.01560627669095993, + 0.004634730517864227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.03367723897099495 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10421334207057953, + 0.04638562723994255, + 0.13244403898715973, + 0.06688954681158066, + 0.28552553057670593, + 0.025599656626582146, + 0.1619425117969513, + 0.06577807664871216, + 0.04888013005256653, + 0.06234158203005791 + ], + "credit_map": [ + 0.014517700299620628, + 0.013501468114554882, + 0.012384769506752491, + 0.010542071424424648, + 0.012858085334300995, + 0.016253774985671043, + 0.013999853283166885, + 0.010155607014894485, + 0.00827164575457573, + 0.004533129278570414, + 0.005089771933853626, + 0.005030158907175064, + 0.006819527596235275, + 0.005330535117536783, + 0.006906269118189812, + 0.004404586274176836, + 0.02817317470908165, + 0.005511742550879717, + 0.009080193936824799, + 0.007573090028017759, + 0.008172262459993362, + 0.008899308741092682, + 0.054100848734378815, + 0.010933409444987774, + 0.00638815201818943, + 0.007462762761861086, + 0.0067565604113042355, + 0.004232415929436684, + 0.013761616311967373, + 0.007696905639022589, + 0.012144613079726696, + 0.008446517400443554, + 0.08283356577157974, + 0.022802358493208885, + 0.014776723459362984, + 0.10636550933122635, + 0.011128575541079044, + 0.01079090777784586, + 0.016359440982341766, + 0.0204684566706419, + 0.002619000617414713, + 0.0030611809343099594, + 0.0017857684288173914, + 0.0019925429951399565, + 0.004329766146838665, + 0.00544970016926527, + 0.0038156325463205576, + 0.0025460647884756327, + 0.02477676421403885, + 0.023168034851551056, + 0.01352632511407137, + 0.034778568893671036, + 0.020330015569925308, + 0.017886720597743988, + 0.014502267353236675, + 0.012973802164196968, + 0.005879556760191917, + 0.007832972332835197, + 0.003796931356191635, + 0.005793819669634104, + 0.0171181820333004, + 0.009714593179523945, + 0.011314702220261097, + 0.004327318165451288, + 0.006834528408944607, + 0.013184260576963425, + 0.0033666007220745087, + 0.003983508329838514, + 0.007887683808803558, + 0.005493697244673967, + 0.004915924277156591, + 0.003213923890143633, + 0.0069949389435350895, + 0.008676102384924889, + 0.00458994135260582, + 0.007888429798185825, + 0.005609030835330486, + 0.01367098093032837, + 0.009416401386260986, + 0.005495752207934856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.009775150567293167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in the whimsical and vibrant style of '90s Disney animation, this static, eye-level medium close-up focuses on a grumpy, scholarly owl. The owl is plump, with meticulously detailed brown and white feathers, and a pair of small, round, gold-rimmed glasses perched precariously on its sharp, dark beak. As he hunches over a thick, dusty tome, his large, expressive golden eyes narrow in comic annoyance, his feathery brows furrowed in deep frustration. The book, bound in deep red leather with a richly textured, worn surface and faded gold foil filigree, rests on a dark wooden surface. With a swift, decisive, and distinctly impatient motion, the owl thrusts its beak downward, pecking fiercely at the book's cover. The impact sends a small, subtle puff of gray dust into the air, where the particles catch the warm, soft light illuminating the scene and swirl for a moment. The setting is a cozy, out-of-focus library, hinted at by the warm tones of the wooden bookshelves. The entire image is defined by crisp, clean hand-drawn lines and a saturated color palette, creating an atmosphere of charm, humor, and focus that perfectly captures the character's scholarly yet cantankerous personality.", + "seed": 3766, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_11/index_8739_seed_3766_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": -0.02214630297385156, + "vlm_logprob": -0.02214630297385156, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.02214630297385156, + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004701630678027868 + ], + "opposite_logprob_tokens": [ + -5.754701614379883 + ], + "logprob_reward": -0.004701630678027868, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004701630678027868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21986354887485504, + 0.07450810074806213, + 0.04537392780184746, + 0.1010451391339302, + 0.04996486380696297, + 0.2292337417602539, + 0.07469171285629272, + 0.07556764781475067, + 0.07980328053236008, + 0.04994801804423332 + ], + "credit_map": [ + 0.022003091871738434, + 0.02086004614830017, + 0.0224814061075449, + 0.03320354223251343, + 0.04086535423994064, + 0.01945546455681324, + 0.02957969345152378, + 0.0314149335026741, + 0.008614521473646164, + 0.00733356736600399, + 0.010840910486876965, + 0.007480968721210957, + 0.011393358930945396, + 0.00960422120988369, + 0.006805588491261005, + 0.01243496872484684, + 0.005572270601987839, + 0.003946287557482719, + 0.005195005796849728, + 0.005663886666297913, + 0.007332054432481527, + 0.005430145654827356, + 0.005683837924152613, + 0.006550436839461327, + 0.009175710380077362, + 0.009799863211810589, + 0.009973948821425438, + 0.008562314324080944, + 0.01984102465212345, + 0.015005450695753098, + 0.010360293090343475, + 0.01832653395831585, + 0.004708587191998959, + 0.005114187952131033, + 0.0049440935254096985, + 0.0034280619584023952, + 0.008226258680224419, + 0.005056531634181738, + 0.006971264723688364, + 0.011515876278281212, + 0.017187772318720818, + 0.0235175471752882, + 0.022716743871569633, + 0.021877102553844452, + 0.045612648129463196, + 0.030169183388352394, + 0.03734941408038139, + 0.03080330416560173, + 0.0073101706802845, + 0.0068521625362336636, + 0.008532258681952953, + 0.012056753039360046, + 0.01241872925311327, + 0.00833707395941019, + 0.009464873000979424, + 0.009719696827232838, + 0.006511225365102291, + 0.008432656526565552, + 0.008345301263034344, + 0.011668414808809757, + 0.013343424536287785, + 0.01182757131755352, + 0.008787669241428375, + 0.0066513861529529095, + 0.0059504336677491665, + 0.005560050718486309, + 0.008970360271632671, + 0.005253170616924763, + 0.015386602841317654, + 0.00597001239657402, + 0.02355506829917431, + 0.009157579392194748, + 0.0034688760060817003, + 0.00440195994451642, + 0.004661548882722855, + 0.0051896777004003525, + 0.009855390526354313, + 0.006444524973630905, + 0.009147558361291885, + 0.00677848095074296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.004701630678027868 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002831143094226718 + ], + "opposite_logprob_tokens": [ + -5.877830982208252 + ], + "logprob_reward": -0.002831143094226718, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002831143094226718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16154153645038605, + 0.11339369416236877, + 0.07393605262041092, + 0.10746864974498749, + 0.05298331752419472, + 0.12729422748088837, + 0.07530318945646286, + 0.18705253303050995, + 0.06743841618299484, + 0.03358836844563484 + ], + "credit_map": [ + 0.01638627238571644, + 0.013909456320106983, + 0.02021290920674801, + 0.0160024706274271, + 0.025561945512890816, + 0.022200508043169975, + 0.018241727724671364, + 0.029026251286268234, + 0.01230600569397211, + 0.010820710100233555, + 0.01736300066113472, + 0.015484915114939213, + 0.016271326690912247, + 0.011963892728090286, + 0.011763134971261024, + 0.017420710995793343, + 0.007612563204020262, + 0.008275434374809265, + 0.00969995092600584, + 0.008081775158643723, + 0.01285326573997736, + 0.008602610789239407, + 0.007571786642074585, + 0.01123866718262434, + 0.00848664902150631, + 0.012220500037074089, + 0.008908659219741821, + 0.012996863573789597, + 0.021572034806013107, + 0.011108310893177986, + 0.0160684771835804, + 0.01610715314745903, + 0.004325434099882841, + 0.007435595151036978, + 0.006063222885131836, + 0.005074384156614542, + 0.009988140314817429, + 0.005124218296259642, + 0.006029894575476646, + 0.008942430838942528, + 0.010218318551778793, + 0.01331609208136797, + 0.014346983283758163, + 0.006976645439863205, + 0.025757642462849617, + 0.017992081120610237, + 0.020827272906899452, + 0.017859188839793205, + 0.007140523754060268, + 0.007479031570255756, + 0.008224476128816605, + 0.012229847721755505, + 0.012755551375448704, + 0.008187337778508663, + 0.009367533959448338, + 0.00991889089345932, + 0.014162421226501465, + 0.018323101103305817, + 0.021074440330266953, + 0.030619382858276367, + 0.03534403815865517, + 0.028153181076049805, + 0.02103271707892418, + 0.018343254923820496, + 0.005086558870971203, + 0.004431006498634815, + 0.006526641547679901, + 0.0051955655217170715, + 0.012856182642281055, + 0.00733244838193059, + 0.01521075889468193, + 0.010799253359436989, + 0.0035013696178793907, + 0.003558460855856538, + 0.0035335347056388855, + 0.004927261266857386, + 0.005889228545129299, + 0.0038873085286468267, + 0.0034873851109296083, + 0.004803820978850126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.002831143094226718 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862820729613304 + ], + "opposite_logprob_tokens": [ + -3.048628091812134 + ], + "logprob_reward": -0.04862820729613304, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862820729613304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13597360253334045, + 0.1329626888036728, + 0.048774100840091705, + 0.035347942262887955, + 0.08427714556455612, + 0.19156447052955627, + 0.046510811895132065, + 0.24968735873699188, + 0.04743532836437225, + 0.027466556057333946 + ], + "credit_map": [ + 0.014633625745773315, + 0.017542995512485504, + 0.017321020364761353, + 0.010164334438741207, + 0.025952383875846863, + 0.015221093781292439, + 0.012043620459735394, + 0.023094523698091507, + 0.011152365244925022, + 0.015513119287788868, + 0.014834059402346611, + 0.008341589011251926, + 0.025908347219228745, + 0.014943198300898075, + 0.0134656373411417, + 0.028804386034607887, + 0.003975315485149622, + 0.004830413032323122, + 0.005837622564285994, + 0.002928219037130475, + 0.00843401812016964, + 0.0061386097222566605, + 0.004404391162097454, + 0.012225514277815819, + 0.00396690471097827, + 0.003403988666832447, + 0.00426727719604969, + 0.0029716913122683764, + 0.004347834270447493, + 0.005203369539231062, + 0.003750384319573641, + 0.0074364906176924706, + 0.005445461720228195, + 0.011563701555132866, + 0.008308406919240952, + 0.006137322634458542, + 0.017889108508825302, + 0.007745741866528988, + 0.010638952255249023, + 0.01654844917356968, + 0.01569933444261551, + 0.02065405622124672, + 0.020773429423570633, + 0.012157234363257885, + 0.03730842098593712, + 0.027808548882603645, + 0.030455267056822777, + 0.026708167046308517, + 0.00421646423637867, + 0.004639685619622469, + 0.005981221329420805, + 0.007777512539178133, + 0.00785214826464653, + 0.0045234644785523415, + 0.005283008329570293, + 0.00623730942606926, + 0.01782524585723877, + 0.022715765982866287, + 0.026694493368268013, + 0.03358533978462219, + 0.05448620766401291, + 0.03803756460547447, + 0.031821344047784805, + 0.024521399289369583, + 0.0036819796077907085, + 0.004157202318310738, + 0.004748465958982706, + 0.0050126127898693085, + 0.009198502637445927, + 0.005470460280776024, + 0.009644449688494205, + 0.005521654617041349, + 0.0026854854077100754, + 0.0028515125159174204, + 0.0028259367682039738, + 0.0035949661396443844, + 0.006092135328799486, + 0.002854295540601015, + 0.0034017693251371384, + 0.0031604543328285217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.04862820729613304 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04862752556800842 + ], + "opposite_logprob_tokens": [ + -3.0486276149749756 + ], + "logprob_reward": -0.04862752556800842, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862752556800842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17161548137664795, + 0.12466710060834885, + 0.04701755940914154, + 0.044847048819065094, + 0.07691796123981476, + 0.1094646006822586, + 0.0439227819442749, + 0.30187085270881653, + 0.05021252855658531, + 0.0294641125947237 + ], + "credit_map": [ + 0.01683051511645317, + 0.023003095760941505, + 0.021814240142703056, + 0.013671314343810081, + 0.03612812981009483, + 0.01889895834028721, + 0.014905408024787903, + 0.026363814249634743, + 0.011168300174176693, + 0.015746913850307465, + 0.015218524262309074, + 0.00936654582619667, + 0.02303144708275795, + 0.0129352156072855, + 0.012711192481219769, + 0.024488961324095726, + 0.004409325309097767, + 0.004835374187678099, + 0.0062157874926924706, + 0.003407705808058381, + 0.007604967337101698, + 0.005556984338909388, + 0.004040664993226528, + 0.010946751572191715, + 0.0042035747319459915, + 0.0046847788617014885, + 0.005182415712624788, + 0.0032391066197305918, + 0.006440052296966314, + 0.00532075110822916, + 0.005231426563113928, + 0.010544943623244762, + 0.005088888108730316, + 0.010797452181577682, + 0.007179083302617073, + 0.006595098879188299, + 0.016249582171440125, + 0.0077895852737128735, + 0.010053976438939571, + 0.013164293952286243, + 0.010163083672523499, + 0.012011853978037834, + 0.014893494546413422, + 0.007101812399923801, + 0.019353462383151054, + 0.015296475030481815, + 0.01615334488451481, + 0.014491063542664051, + 0.003840126097202301, + 0.0038072478491812944, + 0.0045216018334031105, + 0.005557030905038118, + 0.0076257917098701, + 0.004042182583361864, + 0.00793896522372961, + 0.006589836440980434, + 0.021913278847932816, + 0.028559939935803413, + 0.03311796858906746, + 0.04595532640814781, + 0.060791682451963425, + 0.04726892337203026, + 0.03500213846564293, + 0.029261577874422073, + 0.003380850423127413, + 0.003889551619067788, + 0.00433636037632823, + 0.00401885574683547, + 0.010867945849895477, + 0.0054738447070121765, + 0.01077957171946764, + 0.0074655525386333466, + 0.002596460748463869, + 0.003105463460087776, + 0.002982424572110176, + 0.0028588452842086554, + 0.006292376201599836, + 0.002966600703075528, + 0.004475075751543045, + 0.0041868663392961025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.04862752556800842 + }, + { + "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 plump, scholarly owl with brown and white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the owl wear small, round, gold-rimmed glasses on its beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the owl perform a swift motion of thrusting its beak downward to peck the book cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a small puff of gray dust visible swirling in the air after the peck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by crisp, clean hand-drawn lines and a saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943008232861757 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005943008232861757, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943008232861757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3104737102985382, + 0.07411294430494308, + 0.04753783345222473, + 0.07882603257894516, + 0.04171827435493469, + 0.16934062540531158, + 0.04406037554144859, + 0.0805555135011673, + 0.09428859502077103, + 0.059086091816425323 + ], + "credit_map": [ + 0.04010557383298874, + 0.022612180560827255, + 0.038305021822452545, + 0.03470289707183838, + 0.04053367301821709, + 0.03689364716410637, + 0.03425644338130951, + 0.06306428462266922, + 0.008138429373502731, + 0.005561810918152332, + 0.00779699208214879, + 0.006179570220410824, + 0.009049786254763603, + 0.012720618396997452, + 0.007054635789245367, + 0.017611103132367134, + 0.004854495637118816, + 0.00413192855194211, + 0.005505469162017107, + 0.003405763301998377, + 0.007195984944701195, + 0.007280764635652304, + 0.004821246024221182, + 0.010342182591557503, + 0.007021647412329912, + 0.0087728938087821, + 0.00783118512481451, + 0.00701813492923975, + 0.016117850318551064, + 0.009330698288977146, + 0.009797881357371807, + 0.012935739941895008, + 0.0041037932969629765, + 0.0035024050157517195, + 0.004084493033587933, + 0.006860469467937946, + 0.007569231558591127, + 0.006082265637814999, + 0.003920479211956263, + 0.0055951387621462345, + 0.013129530474543571, + 0.016246914863586426, + 0.017172452062368393, + 0.019662799313664436, + 0.031246231868863106, + 0.022379757836461067, + 0.02783959172666073, + 0.021663349121809006, + 0.00552762858569622, + 0.0044080112129449844, + 0.005483884830027819, + 0.00842147134244442, + 0.0055307671427726746, + 0.0049789524637162685, + 0.004674199037253857, + 0.0050354585982859135, + 0.008320387452840805, + 0.011495953425765038, + 0.00639131385833025, + 0.007195319514721632, + 0.02019532211124897, + 0.00898483581840992, + 0.009443084709346294, + 0.008529295213520527, + 0.006913161836564541, + 0.005743441171944141, + 0.009803864173591137, + 0.0064090099185705185, + 0.016577890142798424, + 0.00795944407582283, + 0.029274092987179756, + 0.011607692576944828, + 0.004562155343592167, + 0.005774012766778469, + 0.006504938006401062, + 0.005992637947201729, + 0.010149937123060226, + 0.009116451255977154, + 0.01001262292265892, + 0.0069733355194330215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.005943008232861757 + } + ], + "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/0000167/group_12/index_tensor([1119, 1119])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_tensor([1119, 1119])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..85678742a8647aa7736dc0be67cad5bdf5e9ec9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_tensor([1119, 1119])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shaky, handheld close-up immerses the viewer in a tense moment set in a sunlit forest. The frame is dominated by a woman's hand, knuckles white, muscles tense, gripping a sturdy, gnarled branch, her fingers digging into the intricate, rough bark. The camera tilts slightly, following her subjective experience, as she slowly and deliberately turns her head to the right, her gaze fixed on something off-screen. A wisp of brown hair streaks across her face, her expression resolute—her jaw set, eyes narrowed in intense concentration. Spotted sunlight filters through the tree canopy overhead, creating a dynamic play of light and shadow, highlighting the deep cracks in the bark and the tense lines of her face. The background is a soft-focus blur of lush green foliage. While the implicit sounds of falling leaves and birdsong suggest a tranquil natural atmosphere, her strong grip and alert posture create an overwhelming atmosphere of suspense and anticipation.", + "seed": 30140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_1119_seed_30140_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": -0.02408127912785858, + "vlm_logprob": -0.02408127912785858, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02408127912785858, + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04673749953508377 + ], + "opposite_logprob_tokens": [ + -3.2967374324798584 + ], + "logprob_reward": -0.04673749953508377, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04673749953508377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46057194471359253, + 0.1275336742401123, + 0.0303203035145998, + 0.04719068855047226, + 0.048658546060323715, + 0.03160184994339943, + 0.062145132571458817, + 0.11445582658052444, + 0.034868691116571426, + 0.04265337437391281 + ], + "credit_map": [ + 0.04495416209101677, + 0.02350035309791565, + 0.11842294782400131, + 0.029241057112812996, + 0.020165270194411278, + 0.17661918699741364, + 0.012383216060698032, + 0.03528572991490364, + 0.010909033939242363, + 0.006229579448699951, + 0.009426218457520008, + 0.005131672602146864, + 0.04193698242306709, + 0.04099724069237709, + 0.0071167657151818275, + 0.005786177236586809, + 0.00258915638551116, + 0.0016219386598095298, + 0.00744780246168375, + 0.0019740851130336523, + 0.0024949240032583475, + 0.002529541961848736, + 0.0032477027270942926, + 0.008415150456130505, + 0.005062243901193142, + 0.0024033235386013985, + 0.00984114222228527, + 0.002314767800271511, + 0.0035891833249479532, + 0.004020286723971367, + 0.012032669968903065, + 0.007927070371806622, + 0.004353945143520832, + 0.010629037395119667, + 0.003620339324697852, + 0.010305661708116531, + 0.00740903289988637, + 0.002767325611785054, + 0.004355154931545258, + 0.005218044389039278, + 0.00549004040658474, + 0.004225635901093483, + 0.0028530205599963665, + 0.004040428902953863, + 0.004001996945589781, + 0.0043903556652367115, + 0.0023230037186294794, + 0.0042773662135005, + 0.006240812595933676, + 0.004594156052917242, + 0.004371642135083675, + 0.011926579289138317, + 0.012250917963683605, + 0.008324229158461094, + 0.0062600597739219666, + 0.008176736533641815, + 0.009255615994334221, + 0.007491092197597027, + 0.020829947665333748, + 0.009259677492082119, + 0.006835456006228924, + 0.021152589470148087, + 0.007703310810029507, + 0.031928133219480515, + 0.004092047922313213, + 0.00333107216283679, + 0.006054703611880541, + 0.003440920263528824, + 0.004826562479138374, + 0.0038224058225750923, + 0.005041318945586681, + 0.004259661305695772, + 0.00424862140789628, + 0.002127911662682891, + 0.013240800239145756, + 0.0030877511017024517, + 0.0029649969656020403, + 0.004085995722562075, + 0.007903624325990677, + 0.004993673413991928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04673749953508377 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018207790330052376 + ], + "opposite_logprob_tokens": [ + -4.018208026885986 + ], + "logprob_reward": -0.018207790330052376, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018207790330052376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26832008361816406, + 0.23307400941848755, + 0.10224466025829315, + 0.059332508593797684, + 0.06233764812350273, + 0.009882396087050438, + 0.02301385998725891, + 0.14581134915351868, + 0.04706045985221863, + 0.048923008143901825 + ], + "credit_map": [ + 0.024419404566287994, + 0.0265719722956419, + 0.06695203483104706, + 0.01702135056257248, + 0.01436552032828331, + 0.0996212288737297, + 0.00748883280903101, + 0.011879739351570606, + 0.01854626089334488, + 0.009648950770497322, + 0.01597581058740616, + 0.01014271005988121, + 0.07521679252386093, + 0.07950470596551895, + 0.011274336837232113, + 0.012764453887939453, + 0.011171630583703518, + 0.008344168774783611, + 0.024923108518123627, + 0.009511585347354412, + 0.010705283842980862, + 0.008171437308192253, + 0.00884454045444727, + 0.020572904497385025, + 0.006111480295658112, + 0.009424721822142601, + 0.00997860636562109, + 0.006995304021984339, + 0.007382375653833151, + 0.005861287005245686, + 0.006585035473108292, + 0.006993696093559265, + 0.005928045604377985, + 0.01195184700191021, + 0.007010563276708126, + 0.015084364451467991, + 0.00871989969164133, + 0.003565998049452901, + 0.0034519878681749105, + 0.006624938454478979, + 0.001937770750373602, + 0.0015429885825142264, + 0.0013695418601855636, + 0.0017089091707020998, + 0.001298160175792873, + 0.0008501777774654329, + 0.0006049522780813277, + 0.0005698949680663645, + 0.003915283363312483, + 0.0028902790509164333, + 0.003237818367779255, + 0.002931332914158702, + 0.003294690977782011, + 0.0024927861522883177, + 0.0019159623188897967, + 0.0023357048630714417, + 0.01883346028625965, + 0.008742211386561394, + 0.03562179580330849, + 0.011626355350017548, + 0.007975482381880283, + 0.020706139504909515, + 0.009393525309860706, + 0.03291238844394684, + 0.004291590768843889, + 0.002999798161908984, + 0.003882886376231909, + 0.004303517285734415, + 0.003607611870393157, + 0.010361802764236927, + 0.005257805809378624, + 0.012355447746813297, + 0.0051553212106227875, + 0.0020655905827879906, + 0.0159091018140316, + 0.0027923276647925377, + 0.00246263830922544, + 0.004842135589569807, + 0.009764470160007477, + 0.00593142444267869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.018207790330052376 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11035609245300293, + 0.04545062780380249, + 0.11864527314901352, + 0.04297196865081787, + 0.014919803477823734, + 0.025246424600481987, + 0.06028035283088684, + 0.4049697816371918, + 0.11278623342514038, + 0.06437348574399948 + ], + "credit_map": [ + 0.011454470455646515, + 0.018027398735284805, + 0.028833413496613503, + 0.009119904600083828, + 0.00888153351843357, + 0.00915959756821394, + 0.007874047383666039, + 0.017005713656544685, + 0.00532842380926013, + 0.004831321537494659, + 0.0053696343675255775, + 0.004227497149258852, + 0.009301304817199707, + 0.010486339218914509, + 0.0033867741003632545, + 0.002519326051697135, + 0.011137835681438446, + 0.004300658591091633, + 0.03554597496986389, + 0.007151430007070303, + 0.010007411241531372, + 0.008522856049239635, + 0.00875058677047491, + 0.03322850540280342, + 0.004894573241472244, + 0.007244148291647434, + 0.005731072276830673, + 0.00511019304394722, + 0.004163092002272606, + 0.0027691107243299484, + 0.0067312936298549175, + 0.006328480783849955, + 0.0012524969642981887, + 0.002771825296804309, + 0.000934460258577019, + 0.0027338594663888216, + 0.00304790330119431, + 0.0008721581543795764, + 0.0009568214300088584, + 0.0023502768017351627, + 0.003988452255725861, + 0.0038039851933717728, + 0.0029424671083688736, + 0.0039643640629947186, + 0.0031350713688880205, + 0.0035734735429286957, + 0.0016361279413104057, + 0.0022024803329259157, + 0.007265348453074694, + 0.007207305636256933, + 0.005652157124131918, + 0.01471331249922514, + 0.00391199067234993, + 0.011430774815380573, + 0.004786235745996237, + 0.0053132264874875546, + 0.042222946882247925, + 0.01417853869497776, + 0.0914350152015686, + 0.028219074010849, + 0.01615324057638645, + 0.0757979303598404, + 0.021058466285467148, + 0.11590451747179031, + 0.011158560402691364, + 0.008335985243320465, + 0.011678003706037998, + 0.009266176261007786, + 0.011677373200654984, + 0.022423818707466125, + 0.011896461248397827, + 0.026349849998950958, + 0.004857513587921858, + 0.003730286844074726, + 0.012072897516191006, + 0.006956818979233503, + 0.0061375051736831665, + 0.0056561375968158245, + 0.016797922551631927, + 0.00816439650952816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.03369395062327385 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587678998708725 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587678998708725, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587678998708725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35608986020088196, + 0.2732592523097992, + 0.07694078981876373, + 0.028318393975496292, + 0.022310933098196983, + 0.024107953533530235, + 0.09285038709640503, + 0.052870940417051315, + 0.044851649552583694, + 0.02839977853000164 + ], + "credit_map": [ + 0.023663362488150597, + 0.023995786905288696, + 0.10652106255292892, + 0.02269669622182846, + 0.016103461384773254, + 0.11995821446180344, + 0.01697402633726597, + 0.02617727406322956, + 0.022172836586833, + 0.013891644775867462, + 0.018162395805120468, + 0.01428080815821886, + 0.08691877126693726, + 0.09039493650197983, + 0.013809774070978165, + 0.013628118671476841, + 0.006062016822397709, + 0.003993213642388582, + 0.02074270322918892, + 0.006159273907542229, + 0.006696428172290325, + 0.007299999240785837, + 0.005851869471371174, + 0.020135289058089256, + 0.0025313764344900846, + 0.0032829197589308023, + 0.00382767035625875, + 0.003022103337571025, + 0.0036952514201402664, + 0.0037062920164316893, + 0.003515100572258234, + 0.00473767751827836, + 0.0023769985418766737, + 0.0038624885492026806, + 0.0014006459387019277, + 0.0030559806618839502, + 0.004774891305714846, + 0.0016284582670778036, + 0.0022335511166602373, + 0.00297791650518775, + 0.004715191200375557, + 0.0040514287538826466, + 0.0031315875239670277, + 0.0033896721433848143, + 0.0034102913923561573, + 0.002311880234628916, + 0.0013688154285773635, + 0.0017290846444666386, + 0.01034555397927761, + 0.0077386475168168545, + 0.011415578424930573, + 0.011076350696384907, + 0.019481010735034943, + 0.010817870497703552, + 0.0105585977435112, + 0.01141678262501955, + 0.012039286084473133, + 0.003366724820807576, + 0.01569104939699173, + 0.004440283868461847, + 0.004489402286708355, + 0.003949741832911968, + 0.004637207370251417, + 0.00425724545493722, + 0.0050815981812775135, + 0.003405768657103181, + 0.007621603086590767, + 0.00496671162545681, + 0.005906523205339909, + 0.005237502045929432, + 0.005931182764470577, + 0.00670076347887516, + 0.0014723672065883875, + 0.0013493816368281841, + 0.003409076714888215, + 0.002180845011025667, + 0.002390829147771001, + 0.003849936416372657, + 0.007271177601069212, + 0.006476166192442179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020587678998708725 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.41959139704704285, + 0.06758429110050201, + 0.13061028718948364, + 0.0323113277554512, + 0.02573643997311592, + 0.025751788169145584, + 0.12029694765806198, + 0.06180496886372566, + 0.055108681321144104, + 0.061203889548778534 + ], + "credit_map": [ + 0.052018918097019196, + 0.04708193242549896, + 0.028025271371006966, + 0.029701117426156998, + 0.024645481258630753, + 0.17573370039463043, + 0.013340592384338379, + 0.04904438555240631, + 0.005496887490153313, + 0.004804298747330904, + 0.004383665043860674, + 0.0045852018520236015, + 0.02058529295027256, + 0.02002204954624176, + 0.004512223415076733, + 0.003194669960066676, + 0.013182650320231915, + 0.00435118842869997, + 0.04051303118467331, + 0.004989171400666237, + 0.00948803685605526, + 0.010254212655127048, + 0.009212354198098183, + 0.03861963376402855, + 0.002895437180995941, + 0.0033096347469836473, + 0.005118331406265497, + 0.0023139454424381256, + 0.0020112162455916405, + 0.0027209653053432703, + 0.008846323005855083, + 0.005095476750284433, + 0.002086412627249956, + 0.004224336706101894, + 0.001577912480570376, + 0.0037656270433217287, + 0.0055592358112335205, + 0.0017201339360326529, + 0.0029252320528030396, + 0.003877548035234213, + 0.006285384763032198, + 0.003997963387519121, + 0.003388696350157261, + 0.003324643475934863, + 0.004075117874890566, + 0.001750732772052288, + 0.0016263064462691545, + 0.0013029433321207762, + 0.015441718511283398, + 0.013987241312861443, + 0.011732419021427631, + 0.014931905083358288, + 0.025103485211730003, + 0.013244076631963253, + 0.011936290189623833, + 0.01391981728374958, + 0.01055521983653307, + 0.0037580581847578287, + 0.008852780796587467, + 0.0026913213077932596, + 0.002683223458006978, + 0.012573274783790112, + 0.004755388479679823, + 0.015935702249407768, + 0.005933149252086878, + 0.004040863364934921, + 0.008519441820681095, + 0.0046627274714410305, + 0.006300685461610556, + 0.00842556357383728, + 0.005494360346347094, + 0.01173189003020525, + 0.006160681135952473, + 0.0028250087052583694, + 0.019217589870095253, + 0.00390511192381382, + 0.0038049689028412104, + 0.005969776306301355, + 0.012216025032103062, + 0.007104724645614624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0011794761521741748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld close-up immerses the viewer in a tense moment set in a sunlit forest. The frame is dominated by a woman's hand, knuckles white, muscles tense, gripping a sturdy, gnarled branch, her fingers digging into the intricate, rough bark. The camera tilts slightly, following her subjective experience, as she slowly and deliberately turns her head to the right, her gaze fixed on something off-screen. A wisp of brown hair streaks across her face, her expression resolute—her jaw set, eyes narrowed in intense concentration. Spotted sunlight filters through the tree canopy overhead, creating a dynamic play of light and shadow, highlighting the deep cracks in the bark and the tense lines of her face. The background is a soft-focus blur of lush green foliage. While the implicit sounds of falling leaves and birdsong suggest a tranquil natural atmosphere, her strong grip and alert posture create an overwhelming atmosphere of suspense and anticipation.", + "seed": 30141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_1119_seed_30141_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": -0.04712662072852254, + "vlm_logprob": -0.04712662072852254, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.04712662072852254, + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10491575300693512 + ], + "opposite_logprob_tokens": [ + -2.3549158573150635 + ], + "logprob_reward": -0.10491575300693512, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10491575300693512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04925461485981941, + 0.13063570857048035, + 0.07667422294616699, + 0.09519948810338974, + 0.054907526820898056, + 0.10082972794771194, + 0.1412426382303238, + 0.14557093381881714, + 0.13956019282341003, + 0.06612490862607956 + ], + "credit_map": [ + 0.005871114786714315, + 0.005894929636269808, + 0.004503318108618259, + 0.004536108113825321, + 0.00796180684119463, + 0.007450300268828869, + 0.005985594820231199, + 0.007051446940749884, + 0.01449007447808981, + 0.005652660969644785, + 0.012706161476671696, + 0.006769554223865271, + 0.011715738102793694, + 0.009727522730827332, + 0.02938809059560299, + 0.04018591716885567, + 0.00734413368627429, + 0.009544630534946918, + 0.006559871137142181, + 0.006850525736808777, + 0.011435248889029026, + 0.012794869020581245, + 0.014310862869024277, + 0.00783409085124731, + 0.010895045474171638, + 0.012092125602066517, + 0.014880310744047165, + 0.016424816101789474, + 0.013640684075653553, + 0.00948905386030674, + 0.010856837034225464, + 0.006920620799064636, + 0.008381245657801628, + 0.010497445240616798, + 0.004769391380250454, + 0.005596891511231661, + 0.006750143598765135, + 0.006604471243917942, + 0.00826774537563324, + 0.0040401979349553585, + 0.010745888575911522, + 0.016843624413013458, + 0.010581391863524914, + 0.011404654011130333, + 0.01206088624894619, + 0.017898835241794586, + 0.0137716643512249, + 0.007522794418036938, + 0.007588663604110479, + 0.01896115578711033, + 0.009123323485255241, + 0.01460824441164732, + 0.03822522237896919, + 0.02520737797021866, + 0.017235610634088516, + 0.01029304787516594, + 0.01420691516250372, + 0.017524700611829758, + 0.008811346255242825, + 0.01313081942498684, + 0.022971244528889656, + 0.012441979721188545, + 0.04043764993548393, + 0.01604628935456276, + 0.009378600865602493, + 0.012596691027283669, + 0.013883518055081367, + 0.018752608448266983, + 0.01916903629899025, + 0.027745872735977173, + 0.02091367170214653, + 0.017120206728577614, + 0.0027126125060021877, + 0.010752801783382893, + 0.0019986440893262625, + 0.007818182930350304, + 0.009234852157533169, + 0.00409466540440917, + 0.018225010484457016, + 0.011288151144981384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.10491575300693512 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2720851004123688, + 0.2333039790391922, + 0.04358914494514465, + 0.04198719933629036, + 0.020175425335764885, + 0.03909848630428314, + 0.12818989157676697, + 0.09202443063259125, + 0.09243036061525345, + 0.037115976214408875 + ], + "credit_map": [ + 0.04560908302664757, + 0.026937425136566162, + 0.07474268227815628, + 0.021411176770925522, + 0.025818221271038055, + 0.024487460032105446, + 0.01624993421137333, + 0.036829106509685516, + 0.024004928767681122, + 0.01207837462425232, + 0.03141465038061142, + 0.02423548512160778, + 0.03537379577755928, + 0.015632212162017822, + 0.05553183704614639, + 0.035032715648412704, + 0.004488341975957155, + 0.007196342572569847, + 0.004063256550580263, + 0.003779686288908124, + 0.004675464704632759, + 0.0070027238689363, + 0.008564792573451996, + 0.003818536875769496, + 0.0033044423907995224, + 0.005037542432546616, + 0.004462928976863623, + 0.008250711485743523, + 0.0053042760118842125, + 0.006674895063042641, + 0.005567162297666073, + 0.003385244868695736, + 0.0028362709563225508, + 0.002566732233390212, + 0.002795788925141096, + 0.001611177111044526, + 0.00347963348031044, + 0.0029502541292458773, + 0.002262050984427333, + 0.0016735176322981715, + 0.003879318945109844, + 0.0047815293073654175, + 0.005058209877461195, + 0.005712246987968683, + 0.007254677824676037, + 0.006243895273655653, + 0.004147599451243877, + 0.00202100514434278, + 0.005743824411183596, + 0.013849888928234577, + 0.009134992957115173, + 0.01277115847915411, + 0.03615971654653549, + 0.02640228345990181, + 0.014338823966681957, + 0.009789194911718369, + 0.007821273058652878, + 0.01626153662800789, + 0.009068258106708527, + 0.014256979338824749, + 0.011528770439326763, + 0.008643891662359238, + 0.014273328706622124, + 0.010170387104153633, + 0.005116812419146299, + 0.0063440315425395966, + 0.008235680870711803, + 0.013686936348676682, + 0.011866000480949879, + 0.01927463710308075, + 0.015198438428342342, + 0.012707832269370556, + 0.001868943334557116, + 0.005625818856060505, + 0.001478019985370338, + 0.004098071716725826, + 0.004465537611395121, + 0.002945670625194907, + 0.009968365542590618, + 0.006665547378361225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "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 shot a shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896564900875092 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896564900875092, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896564900875092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06206066161394119, + 0.06567113101482391, + 0.04851546883583069, + 0.026043228805065155, + 0.031037401407957077, + 0.3167576193809509, + 0.11089705675840378, + 0.08813587576150894, + 0.11918436735868454, + 0.13169720768928528 + ], + "credit_map": [ + 0.012113393284380436, + 0.005402391776442528, + 0.011527691036462784, + 0.004983453080058098, + 0.006369825918227434, + 0.007780604064464569, + 0.006737754680216312, + 0.007145542651414871, + 0.008675392717123032, + 0.0034624100662767887, + 0.012947349809110165, + 0.003895181231200695, + 0.008605417795479298, + 0.004769132472574711, + 0.016364354640245438, + 0.006951891351491213, + 0.005281374789774418, + 0.006341582629829645, + 0.004043664783239365, + 0.005155461840331554, + 0.007197564002126455, + 0.008306607604026794, + 0.009476945735514164, + 0.0027122655883431435, + 0.0018972924444824457, + 0.003110273275524378, + 0.001953440485522151, + 0.004579260945320129, + 0.004646449815481901, + 0.002589623676612973, + 0.0043784561567008495, + 0.0028884317725896835, + 0.003972805105149746, + 0.00464742723852396, + 0.003995866514742374, + 0.002634479431435466, + 0.0030288766138255596, + 0.006277449894696474, + 0.003483549924567342, + 0.002996944123879075, + 0.03187502920627594, + 0.06189407408237457, + 0.027231240645051003, + 0.035018835216760635, + 0.043007224798202515, + 0.056004490703344345, + 0.04251673445105553, + 0.019209997728466988, + 0.005914844106882811, + 0.01405308023095131, + 0.012287227436900139, + 0.012566925026476383, + 0.02036193199455738, + 0.016447488218545914, + 0.018229544162750244, + 0.011036010459065437, + 0.006457872688770294, + 0.01514631137251854, + 0.008920935913920403, + 0.010077668353915215, + 0.01204937044531107, + 0.00884732324630022, + 0.013101308606564999, + 0.01353507861495018, + 0.007075697183609009, + 0.006795377004891634, + 0.013647796586155891, + 0.01026199758052826, + 0.015342690981924534, + 0.027317926287651062, + 0.023624498397111893, + 0.015118369832634926, + 0.008262461051344872, + 0.014624527655541897, + 0.006872719153761864, + 0.010637625120580196, + 0.022193677723407745, + 0.007130606099963188, + 0.04146527498960495, + 0.02051031030714512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.07896564900875092 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306614026427269 + ], + "opposite_logprob_tokens": [ + -3.1680660247802734 + ], + "logprob_reward": -0.04306614026427269, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306614026427269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08187492936849594, + 0.13583476841449738, + 0.03748118504881859, + 0.03831329196691513, + 0.08374237269163132, + 0.11540008336305618, + 0.2715698182582855, + 0.11526564508676529, + 0.04066823422908783, + 0.07984978705644608 + ], + "credit_map": [ + 0.012535075657069683, + 0.006256525870412588, + 0.01234306488186121, + 0.0035903463140130043, + 0.008973165415227413, + 0.007792240474373102, + 0.00515333004295826, + 0.025231173262000084, + 0.019321994855999947, + 0.008613198064267635, + 0.036076679825782776, + 0.010910791344940662, + 0.017961816862225533, + 0.0074355462566018105, + 0.019151190295815468, + 0.016363540664315224, + 0.004052831791341305, + 0.0057784258387982845, + 0.0034088611137121916, + 0.00530540244653821, + 0.0050590611062943935, + 0.005858653225004673, + 0.005611102562397718, + 0.0024068430066108704, + 0.003028684761375189, + 0.002804853953421116, + 0.002715960144996643, + 0.008281837217509747, + 0.00440132524818182, + 0.009646879509091377, + 0.004081119783222675, + 0.0033526255283504725, + 0.012504786252975464, + 0.011250118725001812, + 0.00754153449088335, + 0.008240338414907455, + 0.011524374596774578, + 0.013240385800600052, + 0.010547420009970665, + 0.00889340229332447, + 0.012190379202365875, + 0.022573422640562057, + 0.009668758139014244, + 0.013569946400821209, + 0.014309627935290337, + 0.018557369709014893, + 0.016957473009824753, + 0.007573091425001621, + 0.010705678723752499, + 0.027886584401130676, + 0.02027873881161213, + 0.03514169156551361, + 0.07155066728591919, + 0.05316871032118797, + 0.03477449342608452, + 0.018063222989439964, + 0.007980204187333584, + 0.012082968838512897, + 0.010089727118611336, + 0.010892084799706936, + 0.015816431492567062, + 0.015254400670528412, + 0.027995284646749496, + 0.015154536813497543, + 0.002482634037733078, + 0.003962510731071234, + 0.004521722439676523, + 0.005208090879023075, + 0.00753121729940176, + 0.004713182337582111, + 0.0069235218688845634, + 0.005325347650796175, + 0.00361554860137403, + 0.01067274808883667, + 0.0035637628752738237, + 0.009519098326563835, + 0.010910047218203545, + 0.006149863358587027, + 0.021759629249572754, + 0.013659077696502209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.04306614026427269 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010419422760605812 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010419422760605812, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010419422760605812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1439092606306076, + 0.12017598003149033, + 0.052267592400312424, + 0.04727635905146599, + 0.047910355031490326, + 0.07979366928339005, + 0.11959335207939148, + 0.1803850382566452, + 0.14304926991462708, + 0.06563913077116013 + ], + "credit_map": [ + 0.03657190501689911, + 0.011201653629541397, + 0.017916543409228325, + 0.008217396214604378, + 0.012498621828854084, + 0.013572311960160732, + 0.008955033496022224, + 0.0349758081138134, + 0.019899165257811546, + 0.0071482271887362, + 0.019898071885108948, + 0.007417571265250444, + 0.018074316903948784, + 0.011707289144396782, + 0.02129678800702095, + 0.014734544791281223, + 0.0077778035774827, + 0.011043155565857887, + 0.00496964855119586, + 0.006403861567378044, + 0.00657606590539217, + 0.004322624299675226, + 0.005598899908363819, + 0.005575533956289291, + 0.005138925276696682, + 0.004797824192792177, + 0.003769975621253252, + 0.008811772800981998, + 0.008232185617089272, + 0.004441285505890846, + 0.008357824757695198, + 0.003726563649252057, + 0.0067930701188743114, + 0.008404701016843319, + 0.0052765728905797005, + 0.0044330148957669735, + 0.0043919943273067474, + 0.008244875818490982, + 0.0057966457679867744, + 0.004569480661302805, + 0.007991280406713486, + 0.012887292541563511, + 0.011158260516822338, + 0.008583932183682919, + 0.010037959553301334, + 0.014975473284721375, + 0.00855592917650938, + 0.0056035458110272884, + 0.0051756384782493114, + 0.013383535668253899, + 0.012234375812113285, + 0.01090625487267971, + 0.0335126556456089, + 0.02403409406542778, + 0.013582047075033188, + 0.006764747202396393, + 0.022610759362578392, + 0.019553840160369873, + 0.01680500991642475, + 0.015413615852594376, + 0.022261442616581917, + 0.0331839881837368, + 0.030874667689204216, + 0.019681718200445175, + 0.009315206669270992, + 0.011096252128481865, + 0.018389670178294182, + 0.019759617745876312, + 0.018098102882504463, + 0.029676122590899467, + 0.019331226125359535, + 0.017383065074682236, + 0.004897818434983492, + 0.012521548196673393, + 0.003873564535751939, + 0.009203118272125721, + 0.009983914904296398, + 0.005579992197453976, + 0.011101976037025452, + 0.008477198891341686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0010419422760605812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld close-up immerses the viewer in a tense moment set in a sunlit forest. The frame is dominated by a woman's hand, knuckles white, muscles tense, gripping a sturdy, gnarled branch, her fingers digging into the intricate, rough bark. The camera tilts slightly, following her subjective experience, as she slowly and deliberately turns her head to the right, her gaze fixed on something off-screen. A wisp of brown hair streaks across her face, her expression resolute—her jaw set, eyes narrowed in intense concentration. Spotted sunlight filters through the tree canopy overhead, creating a dynamic play of light and shadow, highlighting the deep cracks in the bark and the tense lines of her face. The background is a soft-focus blur of lush green foliage. While the implicit sounds of falling leaves and birdsong suggest a tranquil natural atmosphere, her strong grip and alert posture create an overwhelming atmosphere of suspense and anticipation.", + "seed": 30142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_1119_seed_30142_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": -0.03714994046604261, + "vlm_logprob": -0.03714994046604261, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03714994046604261, + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09572171419858932 + ], + "opposite_logprob_tokens": [ + -2.47072172164917 + ], + "logprob_reward": -0.09572171419858932, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09572171419858932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380695104598999, + 0.08146429806947708, + 0.04486820846796036, + 0.04545581713318825, + 0.3175254166126251, + 0.04864263907074928, + 0.041701603680849075, + 0.1865011304616928, + 0.031313441693782806, + 0.06445786356925964 + ], + "credit_map": [ + 0.026594631373882294, + 0.038286443799734116, + 0.015580541454255581, + 0.020129160955548286, + 0.010304560884833336, + 0.007456794381141663, + 0.012844166718423367, + 0.00687321275472641, + 0.00759466178715229, + 0.011629953049123287, + 0.014783607795834541, + 0.01184014044702053, + 0.007037277799099684, + 0.007371819112449884, + 0.007431504316627979, + 0.013775344006717205, + 0.004672035574913025, + 0.003390649566426873, + 0.008691550232470036, + 0.005633827298879623, + 0.0030644629150629044, + 0.006373088341206312, + 0.004992261063307524, + 0.008050337433815002, + 0.006611105054616928, + 0.004595319740474224, + 0.006329630967229605, + 0.005810718517750502, + 0.003820414887741208, + 0.004341733641922474, + 0.007146605756133795, + 0.0068002925254404545, + 0.021870365366339684, + 0.036447394639253616, + 0.13178975880146027, + 0.022627878934144974, + 0.011669140309095383, + 0.018285490572452545, + 0.02873014286160469, + 0.0461052730679512, + 0.004550578072667122, + 0.0075750392861664295, + 0.00523466058075428, + 0.0066414265893399715, + 0.007927510887384415, + 0.004907173570245504, + 0.00620709965005517, + 0.00559915229678154, + 0.004270093981176615, + 0.007540078833699226, + 0.0033233778085559607, + 0.007384880445897579, + 0.005810298025608063, + 0.0053844028152525425, + 0.003809168003499508, + 0.004179308190941811, + 0.017214413732290268, + 0.016100427135825157, + 0.036182671785354614, + 0.008217384107410908, + 0.01583424210548401, + 0.05602896958589554, + 0.026960201561450958, + 0.009962831623852253, + 0.002463514218106866, + 0.0029070412274450064, + 0.002785108983516693, + 0.0030714606400579214, + 0.004200255498290062, + 0.005904053803533316, + 0.004596579819917679, + 0.005385431461036205, + 0.005622146185487509, + 0.00363441975787282, + 0.003484501503407955, + 0.004102751147001982, + 0.014598107896745205, + 0.009512951597571373, + 0.012410550378262997, + 0.011092441156506538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.09572171419858932 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1743929386138916, + 0.04600748047232628, + 0.02734287455677986, + 0.21130618453025818, + 0.07323287427425385, + 0.03365999460220337, + 0.019915690645575523, + 0.35007569193840027, + 0.026700250804424286, + 0.037366095930337906 + ], + "credit_map": [ + 0.034861840307712555, + 0.04602901637554169, + 0.013291329145431519, + 0.021174060180783272, + 0.01364139188081026, + 0.013194222934544086, + 0.016423460096120834, + 0.01577758975327015, + 0.0068650562316179276, + 0.009265533648431301, + 0.006814100313931704, + 0.0070424070581793785, + 0.0034848961513489485, + 0.003139870474115014, + 0.004236798267811537, + 0.005158810876309872, + 0.0031349665950983763, + 0.002491547027602792, + 0.0041125621646642685, + 0.004156803712248802, + 0.0020836913026869297, + 0.0036755227483808994, + 0.0025956283789128065, + 0.005092148669064045, + 0.020735055208206177, + 0.016092736274003983, + 0.04917645826935768, + 0.017910752445459366, + 0.01540414709597826, + 0.02954607456922531, + 0.026078782975673676, + 0.036362141370773315, + 0.004014660604298115, + 0.004179617390036583, + 0.0029463148675858974, + 0.0035395477898418903, + 0.003866232233121991, + 0.0035937053617089987, + 0.029544882476329803, + 0.02154790237545967, + 0.0029139008838683367, + 0.0019222988048568368, + 0.005492492113262415, + 0.0032300155144184828, + 0.007213952019810677, + 0.003328131977468729, + 0.004391010385006666, + 0.005168191157281399, + 0.0018011496867984533, + 0.0031358269043266773, + 0.0017475842032581568, + 0.0025168072897940874, + 0.0027023993898183107, + 0.003962683025747538, + 0.0018416696693748236, + 0.0022075683809816837, + 0.03048105724155903, + 0.037758197635412216, + 0.0628930926322937, + 0.016875313594937325, + 0.03504037484526634, + 0.09551465511322021, + 0.05327504128217697, + 0.01823793537914753, + 0.0023628193885087967, + 0.0019993074238300323, + 0.004338904749602079, + 0.0024832994677126408, + 0.003514254232868552, + 0.003686907235532999, + 0.003989077173173428, + 0.004325679037719965, + 0.00458850571885705, + 0.0026012978050857782, + 0.0035113992635160685, + 0.003059186739847064, + 0.007853622548282146, + 0.0059342398308217525, + 0.004423031583428383, + 0.005394807551056147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.014215044677257538 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048649102449417114 + ], + "opposite_logprob_tokens": [ + -3.0486490726470947 + ], + "logprob_reward": -0.048649102449417114, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048649102449417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2150958627462387, + 0.07326416671276093, + 0.038134802132844925, + 0.051881369203329086, + 0.11387842893600464, + 0.04811960831284523, + 0.06853768229484558, + 0.07573246955871582, + 0.132236048579216, + 0.18311947584152222 + ], + "credit_map": [ + 0.03171605244278908, + 0.037262652069330215, + 0.0342981331050396, + 0.039138659834861755, + 0.016092944890260696, + 0.013243158347904682, + 0.029090138152241707, + 0.014254160225391388, + 0.01297459565103054, + 0.016880283132195473, + 0.0077148484997451305, + 0.006436251103878021, + 0.006307999137789011, + 0.006883041001856327, + 0.0057303463108837605, + 0.010336804203689098, + 0.005094223190099001, + 0.003494198201224208, + 0.0061653838492929935, + 0.005699852015823126, + 0.0028125729877501726, + 0.006190672051161528, + 0.004172773566097021, + 0.004505131859332323, + 0.007956874556839466, + 0.003529863664880395, + 0.009920991025865078, + 0.0053450264967978, + 0.004786348436027765, + 0.005877318792045116, + 0.006599885411560535, + 0.007865066640079021, + 0.006238595582544804, + 0.012171559035778046, + 0.02345265820622444, + 0.005842089187353849, + 0.005058077163994312, + 0.004823165945708752, + 0.026480238884687424, + 0.029812056571245193, + 0.004378668498247862, + 0.005003591533750296, + 0.008157087489962578, + 0.0054545109160244465, + 0.009863504208624363, + 0.004607472103089094, + 0.00522505771368742, + 0.005429722368717194, + 0.007135939318686724, + 0.01151715312153101, + 0.005367755424231291, + 0.01211561169475317, + 0.009566880762577057, + 0.011266838759183884, + 0.007142644841223955, + 0.0044248709455132484, + 0.004947132430970669, + 0.00928299780935049, + 0.00624434370547533, + 0.005611860193312168, + 0.01135692372918129, + 0.01477225311100483, + 0.01569494977593422, + 0.00782201997935772, + 0.009226012043654919, + 0.00538592878729105, + 0.017494434490799904, + 0.010803885757923126, + 0.015607996843755245, + 0.027306122705340385, + 0.023982565850019455, + 0.02242911048233509, + 0.02697037346661091, + 0.011872407048940659, + 0.02176220342516899, + 0.012225317768752575, + 0.02427348680794239, + 0.038379814475774765, + 0.01940046437084675, + 0.02823542430996895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.048649102449417114 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352116838097572 + ], + "opposite_logprob_tokens": [ + -3.6513521671295166 + ], + "logprob_reward": -0.026352116838097572, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352116838097572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19552607834339142, + 0.026851370930671692, + 0.04620952904224396, + 0.019186142832040787, + 0.5319788455963135, + 0.013272220268845558, + 0.008413285948336124, + 0.09443408250808716, + 0.020136281847953796, + 0.043992042541503906 + ], + "credit_map": [ + 0.02733650803565979, + 0.03536854684352875, + 0.040248803794384, + 0.03590971231460571, + 0.012993838638067245, + 0.0094847921282053, + 0.024056848138570786, + 0.010127049870789051, + 0.003890187479555607, + 0.005633678287267685, + 0.004465591628104448, + 0.0045013041235506535, + 0.0017714535351842642, + 0.0013319506542757154, + 0.0023976971860975027, + 0.00285951211117208, + 0.006324428599327803, + 0.004172695800662041, + 0.011904974468052387, + 0.006869818549603224, + 0.0028341819997876883, + 0.0043166643008589745, + 0.0036306942347437143, + 0.006156079471111298, + 0.002441971329972148, + 0.0022142173256725073, + 0.0037890379317104816, + 0.002300195163115859, + 0.0012882855953648686, + 0.0017114870715886354, + 0.0025539936032146215, + 0.002886958187445998, + 0.025789840146899223, + 0.01610233448445797, + 0.096856988966465, + 0.020058181136846542, + 0.022023949772119522, + 0.020445674657821655, + 0.20606741309165955, + 0.12463457137346268, + 0.0015154120046645403, + 0.001217275857925415, + 0.002422691322863102, + 0.0020415333565324545, + 0.0016695300582796335, + 0.0012756624491885304, + 0.001577466493472457, + 0.001552651752717793, + 0.0006871536606922746, + 0.0014700802275910974, + 0.0006363224238157272, + 0.00179122609551996, + 0.0011415421031415462, + 0.0013792973477393389, + 0.0006606770912185311, + 0.0006469887448474765, + 0.01075395941734314, + 0.0046104546636343, + 0.02869080938398838, + 0.003803699277341366, + 0.004989926237612963, + 0.02918974868953228, + 0.008094264194369316, + 0.0043012406677007675, + 0.0018005846068263054, + 0.0012725949054583907, + 0.0021977811120450497, + 0.0016033626161515713, + 0.0020763452630490065, + 0.0042652105912566185, + 0.0033603268675506115, + 0.0035600787959992886, + 0.008059610612690449, + 0.0033261452335864305, + 0.005539814941585064, + 0.004225424490869045, + 0.00676198722794652, + 0.005930864252150059, + 0.005370863247662783, + 0.004777340684086084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.026352116838097572 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008117241668514907 + ], + "opposite_logprob_tokens": [ + -7.125811576843262 + ], + "logprob_reward": -0.0008117241668514907, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008117241668514907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12756656110286713, + 0.07826919108629227, + 0.03557489439845085, + 0.04429414123296738, + 0.23404818773269653, + 0.027297453954815865, + 0.04405650123953819, + 0.15381747484207153, + 0.043686747550964355, + 0.21138885617256165 + ], + "credit_map": [ + 0.020708303898572922, + 0.0227989349514246, + 0.01261273305863142, + 0.017388051375746727, + 0.01127215102314949, + 0.012681747786700726, + 0.017034180462360382, + 0.01307046227157116, + 0.01820121519267559, + 0.023721951991319656, + 0.007508876733481884, + 0.008459322154521942, + 0.004590149503201246, + 0.0038368136156350374, + 0.005191973876208067, + 0.0067588877864181995, + 0.005198915954679251, + 0.003638163208961487, + 0.008997616358101368, + 0.004084237385541201, + 0.002478670561686158, + 0.003277282463386655, + 0.0033410738687962294, + 0.004558932967483997, + 0.004558320622891188, + 0.0036587915383279324, + 0.008319275453686714, + 0.0049717058427631855, + 0.0034795389510691166, + 0.00543397618457675, + 0.0066437977366149426, + 0.007228734437376261, + 0.010368549264967442, + 0.006847808137536049, + 0.04276595637202263, + 0.008895055390894413, + 0.009580133482813835, + 0.008807167410850525, + 0.09221574664115906, + 0.05456777662038803, + 0.0025850196834653616, + 0.0034445435740053654, + 0.0025494894944131374, + 0.003443207358941436, + 0.004881937988102436, + 0.00358033599331975, + 0.003365370910614729, + 0.003447550581768155, + 0.005666438955813646, + 0.008235003799200058, + 0.004234530031681061, + 0.00825856626033783, + 0.005588008090853691, + 0.0051758443005383015, + 0.003391996957361698, + 0.0035061128437519073, + 0.014704043976962566, + 0.01356804184615612, + 0.02471482753753662, + 0.008077291771769524, + 0.010246898047626019, + 0.047487467527389526, + 0.025527913123369217, + 0.009490999393165112, + 0.0045984662137925625, + 0.003797972807660699, + 0.0054351263679564, + 0.00483607267960906, + 0.005388965364545584, + 0.006039406638592482, + 0.006967037450522184, + 0.006623696070164442, + 0.030314987525343895, + 0.012630208395421505, + 0.023660294711589813, + 0.01479447539895773, + 0.020454872399568558, + 0.0536794476211071, + 0.016548020765185356, + 0.03930655121803284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.0008117241668514907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shaky, handheld close-up immerses the viewer in a tense moment set in a sunlit forest. The frame is dominated by a woman's hand, knuckles white, muscles tense, gripping a sturdy, gnarled branch, her fingers digging into the intricate, rough bark. The camera tilts slightly, following her subjective experience, as she slowly and deliberately turns her head to the right, her gaze fixed on something off-screen. A wisp of brown hair streaks across her face, her expression resolute—her jaw set, eyes narrowed in intense concentration. Spotted sunlight filters through the tree canopy overhead, creating a dynamic play of light and shadow, highlighting the deep cracks in the bark and the tense lines of her face. The background is a soft-focus blur of lush green foliage. While the implicit sounds of falling leaves and birdsong suggest a tranquil natural atmosphere, her strong grip and alert posture create an overwhelming atmosphere of suspense and anticipation.", + "seed": 30143, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_12/index_1119_seed_30143_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": -0.05589077069889754, + "vlm_logprob": -0.05589077069889754, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05589077069889754, + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18723434209823608 + ], + "opposite_logprob_tokens": [ + -1.8122344017028809 + ], + "logprob_reward": -0.18723434209823608, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18723434209823608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14218370616436005, + 0.08145000785589218, + 0.2081710547208786, + 0.0779525488615036, + 0.07446818798780441, + 0.06374724954366684, + 0.07007409632205963, + 0.04555879533290863, + 0.19598959386348724, + 0.040404800325632095 + ], + "credit_map": [ + 0.015305126085877419, + 0.012925702147185802, + 0.03819166123867035, + 0.02077806554734707, + 0.011759171262383461, + 0.02082184888422489, + 0.013304402120411396, + 0.009097722359001637, + 0.007662513293325901, + 0.0056870547123253345, + 0.013514205813407898, + 0.007192566059529781, + 0.007762586232274771, + 0.006498703267425299, + 0.014600053429603577, + 0.018532317131757736, + 0.013809365220367908, + 0.006527322810143232, + 0.05690079554915428, + 0.01288741547614336, + 0.010967308655381203, + 0.010748489759862423, + 0.016559192910790443, + 0.07977116107940674, + 0.009185404516756535, + 0.011122181080281734, + 0.014761988073587418, + 0.00750984949991107, + 0.005445132032036781, + 0.01153162494301796, + 0.0074695744551718235, + 0.010926788672804832, + 0.011566098779439926, + 0.005444632843136787, + 0.01229098904877901, + 0.007730716839432716, + 0.009007570333778858, + 0.010729174129664898, + 0.009999130852520466, + 0.007699870970100164, + 0.005603866185992956, + 0.0093391640111804, + 0.006510581821203232, + 0.006329845171421766, + 0.011709051206707954, + 0.008523392491042614, + 0.007580409292131662, + 0.008150927722454071, + 0.007539212703704834, + 0.007166030816733837, + 0.012290777638554573, + 0.008572169579565525, + 0.009863357990980148, + 0.011318938806653023, + 0.006377268582582474, + 0.006946335081011057, + 0.0032759220339357853, + 0.006860664114356041, + 0.00443664938211441, + 0.00559049379080534, + 0.005336342379450798, + 0.007668789010494947, + 0.0069975294172763824, + 0.005392400547862053, + 0.041071247309446335, + 0.00749411853030324, + 0.014918843284249306, + 0.008641358464956284, + 0.014096839353442192, + 0.03462702035903931, + 0.01408894918859005, + 0.061051204800605774, + 0.003264176892116666, + 0.004303482361137867, + 0.00458155432716012, + 0.007044718135148287, + 0.004513418767601252, + 0.004864311311393976, + 0.00698744086548686, + 0.004845695570111275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.18723434209823608 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596398301422596 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.00596398301422596, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596398301422596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07876314967870712, + 0.03368667885661125, + 0.22431886196136475, + 0.11094355583190918, + 0.06289167702198029, + 0.01685984805226326, + 0.05927446857094765, + 0.024227643385529518, + 0.3235432505607605, + 0.06549084186553955 + ], + "credit_map": [ + 0.00878105591982603, + 0.013045732863247395, + 0.009664277546107769, + 0.00727740116417408, + 0.013052389957010746, + 0.0071217711083590984, + 0.01075825933367014, + 0.009062257595360279, + 0.003702763468027115, + 0.003762598615139723, + 0.004183716140687466, + 0.0036012446507811546, + 0.0034124848898500204, + 0.0024505218025296926, + 0.005328502040356398, + 0.0072448463179171085, + 0.022825665771961212, + 0.013208317570388317, + 0.05623755231499672, + 0.01628706604242325, + 0.011967773549258709, + 0.01149218250066042, + 0.016172530129551888, + 0.076127789914608, + 0.013263300061225891, + 0.014106097631156445, + 0.021979374811053276, + 0.011239570565521717, + 0.009123015217483044, + 0.016498934477567673, + 0.010789125226438046, + 0.013944133184850216, + 0.009624144993722439, + 0.005418157670646906, + 0.01020501833409071, + 0.005392144434154034, + 0.00777810625731945, + 0.009985131211578846, + 0.008955697529017925, + 0.005533276125788689, + 0.0016294107772409916, + 0.002125972416251898, + 0.0015892431838437915, + 0.0020347600802779198, + 0.0036897831596434116, + 0.0023946913424879313, + 0.001442020176909864, + 0.001953966449946165, + 0.0052651818841695786, + 0.005233281292021275, + 0.008299477398395538, + 0.006886949297040701, + 0.009971542283892632, + 0.010722164995968342, + 0.005642757285386324, + 0.007253115531057119, + 0.001951937680132687, + 0.002894954290241003, + 0.0022209316957741976, + 0.00285519752651453, + 0.0030775226186960936, + 0.0047996630892157555, + 0.003351845545694232, + 0.003075591754168272, + 0.07013480365276337, + 0.009669017978012562, + 0.024122050032019615, + 0.00889764167368412, + 0.014466219581663609, + 0.06560713052749634, + 0.016849089413881302, + 0.11379730701446533, + 0.005750712472945452, + 0.00639326823875308, + 0.00894759688526392, + 0.0086686871945858, + 0.0076924595050513744, + 0.007961167953908443, + 0.013588875532150269, + 0.006488073617219925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.00596398301422596 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494222417473793 + ], + "opposite_logprob_tokens": [ + -2.9299421310424805 + ], + "logprob_reward": -0.05494222417473793, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494222417473793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07774786651134491, + 0.04845482483506203, + 0.1918729990720749, + 0.04123549908399582, + 0.03518059104681015, + 0.025785863399505615, + 0.048082660883665085, + 0.061758361756801605, + 0.27746066451072693, + 0.19242067635059357 + ], + "credit_map": [ + 0.009109588339924812, + 0.012394470162689686, + 0.017282353714108467, + 0.007671959698200226, + 0.0076357172802090645, + 0.008553501218557358, + 0.00786114577203989, + 0.007239136379212141, + 0.0038216267712414265, + 0.004227235913276672, + 0.007408671546727419, + 0.005124617367982864, + 0.005941700655966997, + 0.0037102617789059877, + 0.008341341279447079, + 0.009879373013973236, + 0.01466488279402256, + 0.008212283253669739, + 0.050420600920915604, + 0.011227873153984547, + 0.009109193459153175, + 0.009618601761758327, + 0.015474946238100529, + 0.07314459979534149, + 0.004957699682563543, + 0.005500022787600756, + 0.005928605329245329, + 0.0047243800945580006, + 0.005056580062955618, + 0.0066812667064368725, + 0.004820627626031637, + 0.003566314233466983, + 0.00540357967838645, + 0.0036818410735577345, + 0.003909097518771887, + 0.003454594174399972, + 0.004052409436553717, + 0.004653941374272108, + 0.005365903023630381, + 0.004659227095544338, + 0.00271609122864902, + 0.0037872048560529947, + 0.002803680021315813, + 0.00236882152967155, + 0.003646093886345625, + 0.0031728383619338274, + 0.002503645606338978, + 0.004787486512213945, + 0.004256814252585173, + 0.004741885233670473, + 0.006766777951270342, + 0.005836713593453169, + 0.007789636962115765, + 0.008974192664027214, + 0.004255958832800388, + 0.005460677202790976, + 0.005801468621939421, + 0.010680582374334335, + 0.006333769299089909, + 0.009453467093408108, + 0.005194008816033602, + 0.007885470986366272, + 0.008116415701806545, + 0.00829317793250084, + 0.06425941735506058, + 0.012012407183647156, + 0.02372550033032894, + 0.014460332691669464, + 0.016284864395856857, + 0.043907586485147476, + 0.018712332472205162, + 0.0840982049703598, + 0.017397290095686913, + 0.01962892897427082, + 0.025679146870970726, + 0.027574408799409866, + 0.021971171721816063, + 0.02457200363278389, + 0.03576371818780899, + 0.019834011793136597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.05494222417473793 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801562428474426 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.029801562428474426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801562428474426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08820363879203796, + 0.039162203669548035, + 0.5102992057800293, + 0.02987428568303585, + 0.01889001578092575, + 0.013710163533687592, + 0.06573876738548279, + 0.02322613261640072, + 0.15102624893188477, + 0.05986930429935455 + ], + "credit_map": [ + 0.010410296730697155, + 0.011921416036784649, + 0.010383122600615025, + 0.008404376916587353, + 0.012091477401554585, + 0.009972041472792625, + 0.013424153439700603, + 0.01159676257520914, + 0.00475784158334136, + 0.005603312514722347, + 0.004391906782984734, + 0.004163344390690327, + 0.004369090776890516, + 0.0030784569680690765, + 0.005628726910799742, + 0.007169525604695082, + 0.03230467066168785, + 0.01492365077137947, + 0.1438107192516327, + 0.026945430785417557, + 0.024795331060886383, + 0.026344629004597664, + 0.04157436266541481, + 0.1996004432439804, + 0.004139208234846592, + 0.0044059730134904385, + 0.004720856435596943, + 0.003278969554230571, + 0.0034678042866289616, + 0.0039445492438972, + 0.0032094623893499374, + 0.0027074660174548626, + 0.002942149294540286, + 0.0018015530658885837, + 0.0028312793001532555, + 0.0016368296928703785, + 0.0018730394076555967, + 0.003089973935857415, + 0.0026595254894346, + 0.0020556682720780373, + 0.001384078641422093, + 0.002085745567455888, + 0.001772196963429451, + 0.0014799789059907198, + 0.002351351547986269, + 0.001516462303698063, + 0.0014268491650000215, + 0.0016935017192736268, + 0.008036183193325996, + 0.006533666513860226, + 0.012174361385405064, + 0.009729043580591679, + 0.008882302790880203, + 0.008280317299067974, + 0.005982273258268833, + 0.006120624020695686, + 0.001900758477859199, + 0.0035240480210632086, + 0.0020203981548547745, + 0.003130838042125106, + 0.0019178250804543495, + 0.004623940214514732, + 0.0025984481908380985, + 0.003509877948090434, + 0.031518906354904175, + 0.005197926890105009, + 0.012311168946325779, + 0.004839728120714426, + 0.0067232088185846806, + 0.030747059732675552, + 0.008136413060128689, + 0.05155184864997864, + 0.005606831982731819, + 0.005951204802840948, + 0.00814606249332428, + 0.008977800607681274, + 0.006942728068679571, + 0.007579846307635307, + 0.010596586391329765, + 0.0060682473704218864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.029801562428474426 + }, + { + "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 shaky, handheld close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman's hand gripping a sturdy, gnarled branch with white knuckles and tense muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman slowly and deliberately turn her head to the right while her gaze is fixed off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a sunlit forest with spotted sunlight filtering through the canopy creating dynamic light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus blur of lush green foliage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015117417788133025 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015117417788133025, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015117417788133025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1069117933511734, + 0.09139514714479446, + 0.3398488461971283, + 0.06587399542331696, + 0.023976873606443405, + 0.018766434863209724, + 0.07984858751296997, + 0.053872060030698776, + 0.18152327835559845, + 0.037982963025569916 + ], + "credit_map": [ + 0.01641680672764778, + 0.009785955771803856, + 0.01580595038831234, + 0.012646026909351349, + 0.009174860082566738, + 0.014172106049954891, + 0.015234583057463169, + 0.013675500638782978, + 0.009892757050693035, + 0.00803670659661293, + 0.013714505359530449, + 0.009009032510221004, + 0.0099096167832613, + 0.006096606608480215, + 0.016005108132958412, + 0.01873081736266613, + 0.021260419860482216, + 0.010216602124273777, + 0.09651542454957962, + 0.01863873191177845, + 0.015914728865027428, + 0.017354315146803856, + 0.025953391566872597, + 0.13399522006511688, + 0.008375883102416992, + 0.007888582535088062, + 0.009351041167974472, + 0.00648626871407032, + 0.005603177938610315, + 0.011781470850110054, + 0.006482248194515705, + 0.009905329905450344, + 0.0034113104920834303, + 0.0018494799733161926, + 0.002815494081005454, + 0.002377074444666505, + 0.0034546423703432083, + 0.0027248139958828688, + 0.003618505783379078, + 0.0037255515344440937, + 0.0018002546858042479, + 0.002415263094007969, + 0.0017421545926481485, + 0.0025375578552484512, + 0.003848821623250842, + 0.002236920641735196, + 0.0020446141716092825, + 0.002140848897397518, + 0.012918862514197826, + 0.007705193478614092, + 0.01508323848247528, + 0.010889669880270958, + 0.008125286549329758, + 0.007549329660832882, + 0.00973065197467804, + 0.007846360094845295, + 0.0066101280972361565, + 0.012130155228078365, + 0.006844451650977135, + 0.011236161924898624, + 0.0031785278115421534, + 0.004590539261698723, + 0.004659334663301706, + 0.0046227602288126945, + 0.038780756294727325, + 0.00570577522739768, + 0.01706729829311371, + 0.00696959812194109, + 0.008154071867465973, + 0.03417925164103508, + 0.009629525244235992, + 0.06103701516985893, + 0.0032219281420111656, + 0.0029711986426264048, + 0.004559207707643509, + 0.0030918049160391092, + 0.006414816714823246, + 0.006528573110699654, + 0.006576057057827711, + 0.004619377665221691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0015117417788133025 + } + ], + "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/0000167/group_13/index_tensor([7874, 7874])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_tensor([7874, 7874])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f8ed55a53927ff35e9145f06b93121509bed77a8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_tensor([7874, 7874])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, static, panoramic shot transforms the scene into a rock painting, richly grained and mineral-hued. Across a vast wasteland of ochre and dark brown, two giant beetles, the size of armored vehicles, engage in a primal, ferocious struggle. Their obsidian shells gleam with a deep, bronze-metallic sheen, their surfaces crisscrossed with the characteristic rough texture and microscopic cracks of rock painting, as if composed of layers of the purest mineral pigment. One beetle thrusts forward, its six sturdy legs digging into the granular ground for grip, while the other tenses its body, stubbornly resisting the onslaught. The three exaggerated, curved horns atop their heads, like the blades of ancient warriors, are locked in a fierce, silent, and powerful struggle. Each clash unleashes a tremendous force, causing the air around their point of contact to distort and sending fine mineral dust soaring with the violent vibrations. Their movements are imbued with a primal sense of power. One attempts to pry and flip its opponent with its long horns, while the other lowers its body to stabilize its center of gravity. Every muscle contraction is clearly visible, as if one can feel the pulsation within through the rough shell. A dramatic, strong light from low above slants across the scene, carving sharp highlights and deep shadows into their metallic bodies, further intensifying the image's three-dimensionality and rugged texture, and creating an atmosphere of grand, tense, and epic confrontation.", + "seed": 7000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7000_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": -0.021523266099393368, + "vlm_logprob": -0.021523266099393368, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.021523266099393368, + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.03132094070315361 + ], + "opposite_logprob_tokens": [ + -3.5313210487365723 + ], + "logprob_reward": -0.03132094070315361, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03132094070315361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1564226895570755, + 0.08063168823719025, + 0.1206902265548706, + 0.06832364201545715, + 0.1705242097377777, + 0.10310068726539612, + 0.09185076504945755, + 0.1594287008047104, + 0.025437699630856514, + 0.023589694872498512 + ], + "credit_map": [ + 0.01451200246810913, + 0.008390195667743683, + 0.01496847067028284, + 0.06017199531197548, + 0.006571490783244371, + 0.014500065706670284, + 0.010989082045853138, + 0.026319406926631927, + 0.01698935404419899, + 0.004755386617034674, + 0.009692630730569363, + 0.004115683026611805, + 0.005187425762414932, + 0.009078347124159336, + 0.02042897418141365, + 0.010383891873061657, + 0.0399685800075531, + 0.009813123382627964, + 0.03175920248031616, + 0.007019171491265297, + 0.00555714126676321, + 0.009389464743435383, + 0.009585753083229065, + 0.007597793359309435, + 0.008484271354973316, + 0.01103669311851263, + 0.009075763635337353, + 0.007202229928225279, + 0.006904407404363155, + 0.014576783403754234, + 0.005601922050118446, + 0.0054415808990597725, + 0.02012775093317032, + 0.03095710277557373, + 0.020896650850772858, + 0.019775159657001495, + 0.025004232302308083, + 0.02460172213613987, + 0.016994774341583252, + 0.012166837230324745, + 0.013949855230748653, + 0.02216697856783867, + 0.012830741703510284, + 0.011653183959424496, + 0.01618087664246559, + 0.008698005229234695, + 0.009761095978319645, + 0.007859962061047554, + 0.012966248206794262, + 0.013559849932789803, + 0.015023731626570225, + 0.008671985007822514, + 0.014253071509301662, + 0.00992700643837452, + 0.008151445537805557, + 0.009297442622482777, + 0.020230306312441826, + 0.015657061710953712, + 0.02449680306017399, + 0.018057508394122124, + 0.025405729189515114, + 0.019458826631307602, + 0.020902646705508232, + 0.015219823457300663, + 0.0037101113703101873, + 0.003669334575533867, + 0.004095836076885462, + 0.00209961854852736, + 0.0026404974050819874, + 0.0032410232815891504, + 0.0029098265804350376, + 0.003071454120799899, + 0.006494537461549044, + 0.003087111283093691, + 0.006798344198614359, + 0.001900158473290503, + 0.001487676054239273, + 0.001210773247294128, + 0.0013362910831347108, + 0.001274803769774735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.03132094070315361 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.0142021169885993 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.0142021169885993, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142021169885993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37329038977622986, + 0.05401904508471489, + 0.14363951981067657, + 0.06597830355167389, + 0.07562610507011414, + 0.03864671662449837, + 0.08851168304681778, + 0.0386042445898056, + 0.03155910223722458, + 0.09012485295534134 + ], + "credit_map": [ + 0.027700163424015045, + 0.01573256403207779, + 0.034605253487825394, + 0.1484990417957306, + 0.008204085752367973, + 0.030550839379429817, + 0.019719108939170837, + 0.0882793590426445, + 0.010360709391534328, + 0.005092568229883909, + 0.008289546705782413, + 0.005851557943969965, + 0.002576299477368593, + 0.004029251169413328, + 0.009999320842325687, + 0.00781978853046894, + 0.04820684716105461, + 0.009599764831364155, + 0.049555011093616486, + 0.007958498783409595, + 0.00481333676725626, + 0.006977067794650793, + 0.0063453782349824905, + 0.010183618403971195, + 0.009127927012741566, + 0.008925066329538822, + 0.01207763608545065, + 0.006517140660434961, + 0.006693775299936533, + 0.012246954254806042, + 0.00444388622418046, + 0.005945920012891293, + 0.009395347908139229, + 0.011306111700832844, + 0.008356446400284767, + 0.009925228543579578, + 0.010111497715115547, + 0.008652986027300358, + 0.006696504540741444, + 0.011181986890733242, + 0.006104311440140009, + 0.006522431503981352, + 0.005888345651328564, + 0.003869580337777734, + 0.005486744921654463, + 0.004601031541824341, + 0.002789735561236739, + 0.003384535200893879, + 0.014154259115457535, + 0.017838390544056892, + 0.012001068331301212, + 0.008621040731668472, + 0.011020783334970474, + 0.007010228931903839, + 0.008809678256511688, + 0.00905623659491539, + 0.004864789079874754, + 0.004626347217708826, + 0.004724713508039713, + 0.003990361467003822, + 0.007510817609727383, + 0.004783773794770241, + 0.004632594063878059, + 0.0034708494786173105, + 0.00479947030544281, + 0.004350552801042795, + 0.005762640852481127, + 0.0023815506137907505, + 0.0036771600134670734, + 0.003867098828777671, + 0.0029549633618444204, + 0.0037656661588698626, + 0.024774687364697456, + 0.013082953169941902, + 0.025601884350180626, + 0.006263162940740585, + 0.005677338689565659, + 0.00512170884758234, + 0.005410016514360905, + 0.004193099681288004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0142021169885993 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.0337105467915535 + ], + "opposite_logprob_tokens": [ + -3.408710479736328 + ], + "logprob_reward": -0.0337105467915535, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337105467915535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13552197813987732, + 0.33219194412231445, + 0.16749374568462372, + 0.04640312120318413, + 0.11240220069885254, + 0.040355972945690155, + 0.0641222894191742, + 0.03772417828440666, + 0.03865405172109604, + 0.02513049729168415 + ], + "credit_map": [ + 0.010315018706023693, + 0.0059214928187429905, + 0.014313187450170517, + 0.05118353292346001, + 0.0038465410470962524, + 0.010464854538440704, + 0.006608072202652693, + 0.03286927938461304, + 0.06890284270048141, + 0.016679005697369576, + 0.01713968813419342, + 0.017165979370474815, + 0.020938528701663017, + 0.014779816381633282, + 0.1229669526219368, + 0.05361911654472351, + 0.037353985011577606, + 0.012143703177571297, + 0.04877430573105812, + 0.01362475473433733, + 0.008994696661829948, + 0.015250682830810547, + 0.011424102820456028, + 0.01992751844227314, + 0.007025480270385742, + 0.005935680121183395, + 0.007164856884628534, + 0.005331757478415966, + 0.004041746258735657, + 0.009484924376010895, + 0.002992558991536498, + 0.004426116589456797, + 0.01489365566521883, + 0.014886069111526012, + 0.016726166009902954, + 0.012110869400203228, + 0.012881379574537277, + 0.01412420254200697, + 0.010219166986644268, + 0.01656069979071617, + 0.006084592081606388, + 0.006319764070212841, + 0.007543385960161686, + 0.0038456679321825504, + 0.00502177095040679, + 0.005149020813405514, + 0.0027902291622012854, + 0.0036015433724969625, + 0.010882913134992123, + 0.014483796432614326, + 0.008293931372463703, + 0.006680927239358425, + 0.008629955351352692, + 0.0048436434008181095, + 0.005451827310025692, + 0.0048552923835814, + 0.005193354561924934, + 0.003957320004701614, + 0.005119595676660538, + 0.0038877706974744797, + 0.00772575568407774, + 0.004481643438339233, + 0.004223897587507963, + 0.003134839702397585, + 0.005612386390566826, + 0.0054626683704555035, + 0.0066056582145392895, + 0.0034480772446841, + 0.0049324301071465015, + 0.004450637381523848, + 0.003919144626706839, + 0.004223049618303776, + 0.00523557560518384, + 0.004884517285972834, + 0.005880911368876696, + 0.0017743443604558706, + 0.0022005524951964617, + 0.0015934607945382595, + 0.002096716081723571, + 0.0014644176699221134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0337105467915535 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.014190246351063251 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190246351063251, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190246351063251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2553049623966217, + 0.11065885424613953, + 0.08226542174816132, + 0.0408787839114666, + 0.08609937876462936, + 0.061586152762174606, + 0.20800210535526276, + 0.060889389365911484, + 0.05553827062249184, + 0.0387767031788826 + ], + "credit_map": [ + 0.023091014474630356, + 0.012365559116005898, + 0.02582843415439129, + 0.10265568643808365, + 0.007542659994214773, + 0.022550135850906372, + 0.016923896968364716, + 0.04434756562113762, + 0.015562457963824272, + 0.007155255880206823, + 0.014755717478692532, + 0.005243060179054737, + 0.006786522455513477, + 0.008723126724362373, + 0.02946699596941471, + 0.022965719923377037, + 0.011977607384324074, + 0.007256925106048584, + 0.025425013154745102, + 0.006676678545773029, + 0.004298442509025335, + 0.006765035446733236, + 0.007133170031011105, + 0.01273255329579115, + 0.005064764525741339, + 0.00939598772674799, + 0.00506826676428318, + 0.005730802658945322, + 0.00458219088613987, + 0.0047898609191179276, + 0.0028904560022056103, + 0.003356457222253084, + 0.010309315286576748, + 0.014249472878873348, + 0.013115637935698032, + 0.011212694458663464, + 0.011761168949306011, + 0.008968152105808258, + 0.00936770997941494, + 0.00711522251367569, + 0.008700969628989697, + 0.012311762198805809, + 0.009801163338124752, + 0.0056891911663115025, + 0.0077891964465379715, + 0.007294750772416592, + 0.004804997239261866, + 0.0051941219717264175, + 0.033156923949718475, + 0.03345543518662453, + 0.02851199358701706, + 0.015881851315498352, + 0.030460957437753677, + 0.02842053957283497, + 0.01557905413210392, + 0.02253534272313118, + 0.008589443750679493, + 0.005979875568300486, + 0.009690331295132637, + 0.0054842522367835045, + 0.011223972775042057, + 0.0068001775071024895, + 0.007460146676748991, + 0.005661193747073412, + 0.008111688308417797, + 0.008500643074512482, + 0.008443988859653473, + 0.005581970326602459, + 0.0063641746528446674, + 0.006970698945224285, + 0.0060460614040493965, + 0.00551904272288084, + 0.00864511076360941, + 0.003872071160003543, + 0.008140492253005505, + 0.004332179203629494, + 0.0029651967342942953, + 0.003120236098766327, + 0.0032240492291748524, + 0.004477366805076599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.014190246351063251 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.01419247966259718 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.01419247966259718, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419247966259718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42860421538352966, + 0.059104692190885544, + 0.08490075916051865, + 0.0484834760427475, + 0.05876997485756874, + 0.05149518698453903, + 0.07683021575212479, + 0.04955615475773811, + 0.07832945883274078, + 0.06392590701580048 + ], + "credit_map": [ + 0.03278890997171402, + 0.018495937809348106, + 0.03885255753993988, + 0.1672009378671646, + 0.010159728117287159, + 0.03441435843706131, + 0.022436799481511116, + 0.10425497591495514, + 0.00940000731498003, + 0.004335471428930759, + 0.014709758572280407, + 0.003210385562852025, + 0.004670840222388506, + 0.012905987910926342, + 0.003149176947772503, + 0.006723066326230764, + 0.027118725702166557, + 0.007664461154490709, + 0.02004828490316868, + 0.00587126798927784, + 0.004388560075312853, + 0.007738747168332338, + 0.005751150194555521, + 0.0063195619732141495, + 0.007096834480762482, + 0.00791825819760561, + 0.008455604314804077, + 0.005268446635454893, + 0.005073347128927708, + 0.005934807471930981, + 0.0033346691634505987, + 0.005401508416980505, + 0.007998459041118622, + 0.007670903578400612, + 0.00708877295255661, + 0.008127049542963505, + 0.008432483300566673, + 0.007133304141461849, + 0.00516973901540041, + 0.0071492609567940235, + 0.00763199245557189, + 0.01157520990818739, + 0.007801303640007973, + 0.005091907922178507, + 0.007071434520184994, + 0.0037183151580393314, + 0.004621121101081371, + 0.003983901347965002, + 0.012520559132099152, + 0.015354408882558346, + 0.01167191844433546, + 0.006876494735479355, + 0.007921249605715275, + 0.0060878777876496315, + 0.008175773546099663, + 0.008221933618187904, + 0.006373672746121883, + 0.00558222783729434, + 0.006916324608027935, + 0.005027030128985643, + 0.009771780110895634, + 0.005571771878749132, + 0.006163904443383217, + 0.004149445332586765, + 0.011696762405335903, + 0.012273462489247322, + 0.01600712724030018, + 0.005100166890770197, + 0.008389519527554512, + 0.009646252728998661, + 0.0071517666801810265, + 0.008064404129981995, + 0.016932470723986626, + 0.00924929603934288, + 0.01840578392148018, + 0.004238762892782688, + 0.004045410547405481, + 0.0031292615458369255, + 0.00492464704439044, + 0.0030002708081156015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01419247966259718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, static, panoramic shot transforms the scene into a rock painting, richly grained and mineral-hued. Across a vast wasteland of ochre and dark brown, two giant beetles, the size of armored vehicles, engage in a primal, ferocious struggle. Their obsidian shells gleam with a deep, bronze-metallic sheen, their surfaces crisscrossed with the characteristic rough texture and microscopic cracks of rock painting, as if composed of layers of the purest mineral pigment. One beetle thrusts forward, its six sturdy legs digging into the granular ground for grip, while the other tenses its body, stubbornly resisting the onslaught. The three exaggerated, curved horns atop their heads, like the blades of ancient warriors, are locked in a fierce, silent, and powerful struggle. Each clash unleashes a tremendous force, causing the air around their point of contact to distort and sending fine mineral dust soaring with the violent vibrations. Their movements are imbued with a primal sense of power. One attempts to pry and flip its opponent with its long horns, while the other lowers its body to stabilize its center of gravity. Every muscle contraction is clearly visible, as if one can feel the pulsation within through the rough shell. A dramatic, strong light from low above slants across the scene, carving sharp highlights and deep shadows into their metallic bodies, further intensifying the image's three-dimensionality and rugged texture, and creating an atmosphere of grand, tense, and epic confrontation.", + "seed": 7001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7001_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": -0.01412696521729231, + "vlm_logprob": -0.01412696521729231, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01412696521729231, + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.01426604948937893 + ], + "opposite_logprob_tokens": [ + -4.389266014099121 + ], + "logprob_reward": -0.01426604948937893, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01426604948937893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3275163769721985, + 0.12048584967851639, + 0.03821045905351639, + 0.16142049431800842, + 0.02523639239370823, + 0.02652023360133171, + 0.22163377702236176, + 0.019548315554857254, + 0.021293623372912407, + 0.038134437054395676 + ], + "credit_map": [ + 0.10315875709056854, + 0.0150571558624506, + 0.021897634491324425, + 0.01725800894200802, + 0.019439445808529854, + 0.11990499496459961, + 0.015238114632666111, + 0.015562265180051327, + 0.027743006125092506, + 0.01206087414175272, + 0.011551158502697945, + 0.010647510178387165, + 0.009138701483607292, + 0.029299456626176834, + 0.012572534382343292, + 0.007472611032426357, + 0.006137437652796507, + 0.0027703437954187393, + 0.00812454055994749, + 0.004913706332445145, + 0.0032046944834291935, + 0.004034650977700949, + 0.0050261919386684895, + 0.003998894710093737, + 0.02078069932758808, + 0.019268829375505447, + 0.023736387491226196, + 0.0151003897190094, + 0.0315750315785408, + 0.0175118800252676, + 0.017524434253573418, + 0.015922848135232925, + 0.004356893710792065, + 0.002158722374588251, + 0.007005035411566496, + 0.0020907579455524683, + 0.0026131870690733194, + 0.0030449749901890755, + 0.0017724712379276752, + 0.002194348955526948, + 0.0030909166671335697, + 0.004310202784836292, + 0.002884285757318139, + 0.0025568543933331966, + 0.004241904243826866, + 0.0034103563521057367, + 0.0036570036318153143, + 0.002368710469454527, + 0.019916057586669922, + 0.0064759645611047745, + 0.08083486557006836, + 0.011156018823385239, + 0.004453870467841625, + 0.08946499973535538, + 0.004364443477243185, + 0.004967567510902882, + 0.003179041435942054, + 0.0020557111129164696, + 0.005209028720855713, + 0.0016170351300388575, + 0.0017209658399224281, + 0.0018589942483231425, + 0.001418228610418737, + 0.002489311620593071, + 0.002939091995358467, + 0.003218380268663168, + 0.003170850919559598, + 0.0021163939964026213, + 0.0026020612567663193, + 0.0018534798873588443, + 0.002157820388674736, + 0.0032355445437133312, + 0.005154064856469631, + 0.005878081079572439, + 0.005682723131030798, + 0.0047775921411812305, + 0.0045852381736040115, + 0.0046214512549340725, + 0.0031745019368827343, + 0.004260786343365908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.01426604948937893 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.014203997328877449 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203997328877449, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203997328877449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5200212597846985, + 0.07773242890834808, + 0.023849740624427795, + 0.04357311874628067, + 0.029338039457798004, + 0.034148216247558594, + 0.07714800536632538, + 0.0303023774176836, + 0.07814516127109528, + 0.08574165403842926 + ], + "credit_map": [ + 0.17981216311454773, + 0.02191723696887493, + 0.0332244448363781, + 0.022750675678253174, + 0.01756403036415577, + 0.2076655477285385, + 0.012382511980831623, + 0.024704640731215477, + 0.021588055416941643, + 0.004216583911329508, + 0.008163878694176674, + 0.004018215928226709, + 0.0045119295828044415, + 0.02549983561038971, + 0.005500560626387596, + 0.004233369138091803, + 0.003588082268834114, + 0.001654555555433035, + 0.003920432645827532, + 0.003343047806993127, + 0.0020869383588433266, + 0.002968195825815201, + 0.003106169169768691, + 0.003182319924235344, + 0.007031711749732494, + 0.0034253508783876896, + 0.007625875063240528, + 0.004132229834794998, + 0.0050334930419921875, + 0.006082504987716675, + 0.004696979653090239, + 0.005544975865632296, + 0.005708649288862944, + 0.002089279005303979, + 0.00962953083217144, + 0.003073458094149828, + 0.0021633782889693975, + 0.0019184596603736281, + 0.002465058583766222, + 0.002290224190801382, + 0.007600411307066679, + 0.0038188069593161345, + 0.003629089333117008, + 0.002372392686083913, + 0.0029045897535979748, + 0.0075098066590726376, + 0.003343974705785513, + 0.002969142748042941, + 0.007176960818469524, + 0.0019244819413870573, + 0.01521157193928957, + 0.0029823363292962313, + 0.002014381578192115, + 0.032341063022613525, + 0.0028544028755277395, + 0.012642808258533478, + 0.0049099004827439785, + 0.0030622121412307024, + 0.005287692882120609, + 0.003163906978443265, + 0.003384374314919114, + 0.0035964236594736576, + 0.0024901425931602716, + 0.004407722502946854, + 0.013129687868058681, + 0.007021002471446991, + 0.019952241331338882, + 0.006244364194571972, + 0.0048086978495121, + 0.007688543293625116, + 0.006205623969435692, + 0.013094998896121979, + 0.014378744177520275, + 0.012292417697608471, + 0.01693311333656311, + 0.00878225639462471, + 0.008786817081272602, + 0.00874735601246357, + 0.004914155229926109, + 0.01090678945183754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.014203997328877449 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.018198074772953987 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018198074772953987, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198074772953987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34823840856552124, + 0.057252682745456696, + 0.029252756386995316, + 0.07613915950059891, + 0.03452160581946373, + 0.03420231118798256, + 0.2471962422132492, + 0.044643815606832504, + 0.08628279715776443, + 0.0422702431678772 + ], + "credit_map": [ + 0.11925245076417923, + 0.014830179512500763, + 0.023141833022236824, + 0.016667872667312622, + 0.012280954048037529, + 0.1358509659767151, + 0.008579387329518795, + 0.01763477362692356, + 0.016227006912231445, + 0.003559476463124156, + 0.005716771353036165, + 0.0034189496655017138, + 0.0038738935254514217, + 0.017584875226020813, + 0.00413625780493021, + 0.002735449466854334, + 0.0046036685816943645, + 0.0020017551723867655, + 0.005706556141376495, + 0.003805038519203663, + 0.002868174808099866, + 0.004221166018396616, + 0.0025451150722801685, + 0.0035012811422348022, + 0.010076707229018211, + 0.008445599116384983, + 0.011363530531525612, + 0.007182032335549593, + 0.013095014728605747, + 0.010181630961596966, + 0.008037440478801727, + 0.007757203187793493, + 0.006223446223884821, + 0.0026525852736085653, + 0.011081797070801258, + 0.0036551461089402437, + 0.002581304172053933, + 0.003553812624886632, + 0.00213220133446157, + 0.0026413125451654196, + 0.004999908152967691, + 0.004382947459816933, + 0.003467835485935211, + 0.0034551064018160105, + 0.005823078565299511, + 0.003890948137268424, + 0.004481777083128691, + 0.0037007108330726624, + 0.021385326981544495, + 0.005921243689954281, + 0.07126511633396149, + 0.01061766967177391, + 0.006174110807478428, + 0.10378210246562958, + 0.006319985259324312, + 0.021730680018663406, + 0.0077513596042990685, + 0.004638653248548508, + 0.006229932885617018, + 0.0040494659915566444, + 0.005487494170665741, + 0.003987525589764118, + 0.004083957988768816, + 0.00841542612761259, + 0.013768678531050682, + 0.007172060199081898, + 0.023525431752204895, + 0.007134482730180025, + 0.0055299135856330395, + 0.008595461025834084, + 0.006566574331372976, + 0.013990195468068123, + 0.006926323752850294, + 0.006664653774350882, + 0.005693694110959768, + 0.004490560386329889, + 0.005796870216727257, + 0.005305393133312464, + 0.0025571302976459265, + 0.004835615865886211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.018198074772953987 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.014188366010785103 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188366010785103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188366010785103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2177940160036087, + 0.268696129322052, + 0.05839015543460846, + 0.08219047635793686, + 0.031928982585668564, + 0.0645289421081543, + 0.15486279129981995, + 0.02139713056385517, + 0.05390308052301407, + 0.0463082455098629 + ], + "credit_map": [ + 0.07347751408815384, + 0.010762094520032406, + 0.013340447098016739, + 0.010041443631052971, + 0.009437162429094315, + 0.08449947834014893, + 0.006360554601997137, + 0.009875321760773659, + 0.08788511157035828, + 0.012117788195610046, + 0.026900561526417732, + 0.010256685316562653, + 0.01389111764729023, + 0.0975717306137085, + 0.010596844367682934, + 0.00947631150484085, + 0.009735839441418648, + 0.0037652452010661364, + 0.01654207333922386, + 0.006235015578567982, + 0.00552546139806509, + 0.005855864845216274, + 0.006485779769718647, + 0.004244877025485039, + 0.012998784892261028, + 0.008667691610753536, + 0.012867823243141174, + 0.006492855492979288, + 0.010332836769521236, + 0.013873319141566753, + 0.008610673248767853, + 0.008346488699316978, + 0.004169887397438288, + 0.0034590447321534157, + 0.005105030257254839, + 0.0038715549744665623, + 0.00429165456444025, + 0.003938525915145874, + 0.003218586090952158, + 0.003874698420986533, + 0.00909474864602089, + 0.010021684691309929, + 0.007882505655288696, + 0.006074478384107351, + 0.007983422838151455, + 0.005730935372412205, + 0.012214159592986107, + 0.005527009721845388, + 0.013562249951064587, + 0.0033870707266032696, + 0.040537454187870026, + 0.0064716399647295475, + 0.0033751598093658686, + 0.06584019958972931, + 0.004240749403834343, + 0.017448274418711662, + 0.00257497769780457, + 0.002801267197355628, + 0.0029178985860198736, + 0.0024161702021956444, + 0.003141478169709444, + 0.0026308330707252026, + 0.002084792358800769, + 0.002829713746905327, + 0.007669474929571152, + 0.006239403039216995, + 0.009715232066810131, + 0.005486201494932175, + 0.004404047038406134, + 0.0076963878236711025, + 0.0052978103049099445, + 0.007394523825496435, + 0.006878898944705725, + 0.010156293399631977, + 0.007294633891433477, + 0.005235226824879646, + 0.003732955316081643, + 0.004599852953106165, + 0.0037246858701109886, + 0.004685696214437485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.014188366010785103 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.009778338484466076 + ], + "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": [ + 0.44045668840408325, + 0.042020052671432495, + 0.02742321416735649, + 0.03456323221325874, + 0.05265609920024872, + 0.03973174840211868, + 0.22755350172519684, + 0.021805331110954285, + 0.05039236322045326, + 0.06339780241250992 + ], + "credit_map": [ + 0.15750733017921448, + 0.0156550295650959, + 0.0267434474080801, + 0.017059043049812317, + 0.015223732218146324, + 0.18136778473854065, + 0.008918559178709984, + 0.01798177883028984, + 0.009216800332069397, + 0.0036985811311751604, + 0.004618192091584206, + 0.004077737219631672, + 0.0035680984146893024, + 0.009307409636676311, + 0.004681435879319906, + 0.002851794706657529, + 0.004154661204665899, + 0.0014889286831021309, + 0.005558987148106098, + 0.0044877855107188225, + 0.0016288721235468984, + 0.002025228925049305, + 0.004267892800271511, + 0.0038108567241579294, + 0.0054169888608157635, + 0.004429982043802738, + 0.004558088723570108, + 0.0027594752609729767, + 0.005525527521967888, + 0.005587344523519278, + 0.0027544766198843718, + 0.0035313477274030447, + 0.010527340695261955, + 0.003647151170298457, + 0.01934739574790001, + 0.005571957211941481, + 0.0024416749365627766, + 0.003317029681056738, + 0.0034845927730202675, + 0.004318954888731241, + 0.009097649715840816, + 0.004463472403585911, + 0.004853658378124237, + 0.0030220348853617907, + 0.005052834749221802, + 0.005482333246618509, + 0.004338893573731184, + 0.0034208670258522034, + 0.02078857272863388, + 0.005834309384226799, + 0.06929805874824524, + 0.010048759169876575, + 0.00456102192401886, + 0.09555577486753464, + 0.004658103454858065, + 0.016808900982141495, + 0.003911416046321392, + 0.0022151884622871876, + 0.003451986936852336, + 0.001990176271647215, + 0.002835535444319248, + 0.0029143341816961765, + 0.0015413141809403896, + 0.0029453798197209835, + 0.007393764331936836, + 0.0071884868666529655, + 0.009919079020619392, + 0.0058952877297997475, + 0.0037520499899983406, + 0.0060234712436795235, + 0.004301869310438633, + 0.005918354261666536, + 0.009696717374026775, + 0.008565597236156464, + 0.010644491761922836, + 0.006737645715475082, + 0.0057632955722510815, + 0.008463137783110142, + 0.004306873306632042, + 0.009220048785209656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.009778338484466076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, static, panoramic shot transforms the scene into a rock painting, richly grained and mineral-hued. Across a vast wasteland of ochre and dark brown, two giant beetles, the size of armored vehicles, engage in a primal, ferocious struggle. Their obsidian shells gleam with a deep, bronze-metallic sheen, their surfaces crisscrossed with the characteristic rough texture and microscopic cracks of rock painting, as if composed of layers of the purest mineral pigment. One beetle thrusts forward, its six sturdy legs digging into the granular ground for grip, while the other tenses its body, stubbornly resisting the onslaught. The three exaggerated, curved horns atop their heads, like the blades of ancient warriors, are locked in a fierce, silent, and powerful struggle. Each clash unleashes a tremendous force, causing the air around their point of contact to distort and sending fine mineral dust soaring with the violent vibrations. Their movements are imbued with a primal sense of power. One attempts to pry and flip its opponent with its long horns, while the other lowers its body to stabilize its center of gravity. Every muscle contraction is clearly visible, as if one can feel the pulsation within through the rough shell. A dramatic, strong light from low above slants across the scene, carving sharp highlights and deep shadows into their metallic bodies, further intensifying the image's three-dimensionality and rugged texture, and creating an atmosphere of grand, tense, and epic confrontation.", + "seed": 7002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7002_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": -0.029039162583649157, + "vlm_logprob": -0.029039162583649157, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.029039162583649157, + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.0637691542506218 + ], + "opposite_logprob_tokens": [ + -2.8137691020965576 + ], + "logprob_reward": -0.0637691542506218, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0637691542506218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20608016848564148, + 0.0922129675745964, + 0.07346033304929733, + 0.05017034709453583, + 0.05713002383708954, + 0.07731621712446213, + 0.09891170263290405, + 0.05303588882088661, + 0.05673122778534889, + 0.23495109379291534 + ], + "credit_map": [ + 0.03314590081572533, + 0.024670470505952835, + 0.03987095504999161, + 0.021831011399626732, + 0.01848527416586876, + 0.028919639065861702, + 0.01647707261145115, + 0.02267986349761486, + 0.012974108569324017, + 0.015564870089292526, + 0.013607031665742397, + 0.009356023743748665, + 0.009689311496913433, + 0.013857276178896427, + 0.008402962237596512, + 0.008761378936469555, + 0.010353677906095982, + 0.009579292498528957, + 0.008973313495516777, + 0.006677363533526659, + 0.011775261722505093, + 0.011053645983338356, + 0.008014538325369358, + 0.007033236790448427, + 0.007061677053570747, + 0.005100807640701532, + 0.009145254269242287, + 0.005553551949560642, + 0.006362561136484146, + 0.006560866720974445, + 0.003854628186672926, + 0.006530997343361378, + 0.006991012487560511, + 0.007364283315837383, + 0.00725797563791275, + 0.008946452289819717, + 0.005111024249345064, + 0.007292791735380888, + 0.006215967237949371, + 0.007950520142912865, + 0.01329793781042099, + 0.009763737209141254, + 0.012306278571486473, + 0.011172502301633358, + 0.006212548352777958, + 0.009319991804659367, + 0.006772504188120365, + 0.008470720611512661, + 0.01897839643061161, + 0.011494278907775879, + 0.01669338531792164, + 0.011046458035707474, + 0.00862304586917162, + 0.012682954780757427, + 0.008304139599204063, + 0.011089050211012363, + 0.010971661657094955, + 0.004214255604892969, + 0.009798355400562286, + 0.0052667465060949326, + 0.0036676230374723673, + 0.003853156929835677, + 0.006427563726902008, + 0.008836526423692703, + 0.008653358556330204, + 0.005706084426492453, + 0.00938648171722889, + 0.004827591590583324, + 0.005710041616111994, + 0.004529872443526983, + 0.011264348402619362, + 0.006653450429439545, + 0.03961486741900444, + 0.024857737123966217, + 0.031059660017490387, + 0.02178216725587845, + 0.02072184905409813, + 0.030390454456210136, + 0.03165162727236748, + 0.03487272933125496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0637691542506218 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.016074584797024727 + ], + "opposite_logprob_tokens": [ + -4.1410746574401855 + ], + "logprob_reward": -0.016074584797024727, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074584797024727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2904527187347412, + 0.10956916213035583, + 0.06955777108669281, + 0.07832124084234238, + 0.06956050544977188, + 0.09424414485692978, + 0.09286116808652878, + 0.07818127423524857, + 0.04110173508524895, + 0.07615023851394653 + ], + "credit_map": [ + 0.049661099910736084, + 0.03127783536911011, + 0.04524307698011398, + 0.03598618507385254, + 0.0293444711714983, + 0.046308767050504684, + 0.02275923453271389, + 0.029872054234147072, + 0.019445816054940224, + 0.01265624724328518, + 0.014707166701555252, + 0.010685845278203487, + 0.010203992016613483, + 0.01587522029876709, + 0.011430252343416214, + 0.014564629644155502, + 0.013707649894058704, + 0.007385503966361284, + 0.010492785833775997, + 0.006031885743141174, + 0.005394482519477606, + 0.010154083371162415, + 0.006591959856450558, + 0.009799416176974773, + 0.01258979644626379, + 0.007998154498636723, + 0.013350052759051323, + 0.011960514821112156, + 0.006742320489138365, + 0.010367396287620068, + 0.00665070116519928, + 0.008662311360239983, + 0.015087013132870197, + 0.008710583671927452, + 0.012473239563405514, + 0.006891487631946802, + 0.004140842705965042, + 0.007335959933698177, + 0.004204775206744671, + 0.010716600343585014, + 0.018697932362556458, + 0.009775337763130665, + 0.021569548174738884, + 0.011196616105735302, + 0.007191901560872793, + 0.008257271721959114, + 0.008321343921124935, + 0.009234190918505192, + 0.01873324066400528, + 0.010384676977992058, + 0.023602118715643883, + 0.008307556621730328, + 0.00760603416711092, + 0.0054956772364676, + 0.00890378188341856, + 0.009828083217144012, + 0.01384778879582882, + 0.007813304662704468, + 0.013575233519077301, + 0.009562762454152107, + 0.005369482561945915, + 0.006193514447659254, + 0.008781344629824162, + 0.0130378482863307, + 0.007399001624435186, + 0.004732586909085512, + 0.00970652885735035, + 0.004273902624845505, + 0.003511515213176608, + 0.0028230706229805946, + 0.005590174347162247, + 0.0030649553518742323, + 0.011036509647965431, + 0.010338212363421917, + 0.009782812558114529, + 0.010043411515653133, + 0.005611145403236151, + 0.008635614067316055, + 0.009447132237255573, + 0.01125540304929018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016074584797024727 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2025613933801651, + 0.07295197993516922, + 0.11888635903596878, + 0.052285335958004, + 0.09122925251722336, + 0.12863613665103912, + 0.08269219100475311, + 0.04308932647109032, + 0.12671199440956116, + 0.0809561014175415 + ], + "credit_map": [ + 0.027012338861823082, + 0.02495863102376461, + 0.030100233852863312, + 0.02706197276711464, + 0.026555029675364494, + 0.030864344909787178, + 0.01754053123295307, + 0.01846831850707531, + 0.012483618222177029, + 0.009472852572798729, + 0.009592213667929173, + 0.0072552538476884365, + 0.011185516603291035, + 0.010384771041572094, + 0.0063005294650793076, + 0.006277225911617279, + 0.014953581616282463, + 0.014302287250757217, + 0.013786065392196178, + 0.011627072468400002, + 0.02008637599647045, + 0.01590942218899727, + 0.014120037667453289, + 0.014101509004831314, + 0.008656357415020466, + 0.0054528065957129, + 0.010647446848452091, + 0.007381889503449202, + 0.00584699772298336, + 0.0053044031374156475, + 0.0031863516196608543, + 0.005809085443615913, + 0.016431275755167007, + 0.012332947924733162, + 0.014327261596918106, + 0.011657394468784332, + 0.00757720647379756, + 0.008875546045601368, + 0.00835301075130701, + 0.0116746099665761, + 0.03297781944274902, + 0.009346621111035347, + 0.034963227808475494, + 0.01050893310457468, + 0.006196817848831415, + 0.011285126209259033, + 0.011881368234753609, + 0.011476232670247555, + 0.01751355268061161, + 0.006709934677928686, + 0.015410888940095901, + 0.007802004460245371, + 0.006792904809117317, + 0.008176933974027634, + 0.009685507044196129, + 0.01060046162456274, + 0.008711420930922031, + 0.0044136447831988335, + 0.0095357121899724, + 0.00444830721244216, + 0.0033105278853327036, + 0.0037324761506170034, + 0.004058267921209335, + 0.0048789698630571365, + 0.01961050182580948, + 0.020761655643582344, + 0.018400656059384346, + 0.016000550240278244, + 0.011143469251692295, + 0.010526888072490692, + 0.017938945442438126, + 0.012329318560659885, + 0.010779831558465958, + 0.012210036627948284, + 0.011153246276080608, + 0.012176458723843098, + 0.006874218117445707, + 0.009550136514008045, + 0.00754462368786335, + 0.010667557828128338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.020591532811522484 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.033689454197883606 + ], + "opposite_logprob_tokens": [ + -3.408689498901367 + ], + "logprob_reward": -0.033689454197883606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033689454197883606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1134614497423172, + 0.11945347487926483, + 0.026239808648824692, + 0.050830330699682236, + 0.0825534239411354, + 0.12472622841596603, + 0.04280262440443039, + 0.09716543555259705, + 0.0987963005900383, + 0.24397099018096924 + ], + "credit_map": [ + 0.025199593976140022, + 0.008970812894403934, + 0.01989244855940342, + 0.012296771630644798, + 0.013061366975307465, + 0.013075536116957664, + 0.010273118503391743, + 0.01069179829210043, + 0.01819358579814434, + 0.02004944533109665, + 0.020521709695458412, + 0.010443340986967087, + 0.011568349786102772, + 0.018211806192994118, + 0.01127026230096817, + 0.009194977581501007, + 0.0041674962267279625, + 0.002871154807507992, + 0.0038261034060269594, + 0.0030858030077069998, + 0.00241777952760458, + 0.0028630425222218037, + 0.00311791873537004, + 0.0038905092515051365, + 0.006740476470440626, + 0.006647862493991852, + 0.009767052717506886, + 0.006493312772363424, + 0.005111247766762972, + 0.008182904683053493, + 0.0032528808806091547, + 0.0046345884911715984, + 0.011159670539200306, + 0.011929881758987904, + 0.011710820719599724, + 0.009193015284836292, + 0.008394713513553143, + 0.012155965901911259, + 0.007528399582952261, + 0.010480955243110657, + 0.019796133041381836, + 0.014457253739237785, + 0.017051927745342255, + 0.018315382301807404, + 0.009961698204278946, + 0.017740851268172264, + 0.012882995419204235, + 0.014519981108605862, + 0.008269719779491425, + 0.006222350522875786, + 0.008286085911095142, + 0.0037994179874658585, + 0.0036577987484633923, + 0.004250475205481052, + 0.0035654185339808464, + 0.004751351196318865, + 0.016344018280506134, + 0.012302007526159286, + 0.01575361378490925, + 0.0127192297950387, + 0.005948561243712902, + 0.009105842560529709, + 0.00740596791729331, + 0.017586179077625275, + 0.01512459572404623, + 0.01470431312918663, + 0.014602906070649624, + 0.012003561481833458, + 0.011220846325159073, + 0.009087491780519485, + 0.012486491352319717, + 0.009566091932356358, + 0.036593589931726456, + 0.03145504370331764, + 0.03158915042877197, + 0.03485196828842163, + 0.024726564064621925, + 0.03590814769268036, + 0.02395128458738327, + 0.024895213544368744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.033689454197883606 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.01107108686119318 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.01107108686119318, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107108686119318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19790585339069366, + 0.10760503262281418, + 0.03118518739938736, + 0.03721019998192787, + 0.032205186784267426, + 0.2959035038948059, + 0.07891551405191422, + 0.047203924506902695, + 0.11924441158771515, + 0.05262119323015213 + ], + "credit_map": [ + 0.03342363238334656, + 0.02003372274339199, + 0.03312680870294571, + 0.024522896856069565, + 0.01987435296177864, + 0.030090175569057465, + 0.015974877402186394, + 0.020859388634562492, + 0.015988921746611595, + 0.01202103029936552, + 0.01435114350169897, + 0.012412600219249725, + 0.010518818162381649, + 0.013916504569351673, + 0.01458756998181343, + 0.013808457180857658, + 0.006624745205044746, + 0.0039830864407122135, + 0.0044425674714148045, + 0.002126311184838414, + 0.003021710552275181, + 0.004622635897248983, + 0.002312205731868744, + 0.004051927477121353, + 0.007375007960945368, + 0.0027042871806770563, + 0.006928589195013046, + 0.0036476165987551212, + 0.005127660930156708, + 0.005274991504848003, + 0.0026395239401608706, + 0.0035125252325087786, + 0.005505464505404234, + 0.004553015343844891, + 0.004724110011011362, + 0.003667611163109541, + 0.0031594624742865562, + 0.004818539600819349, + 0.0025331482756882906, + 0.0032438370399177074, + 0.08587890863418579, + 0.03718389570713043, + 0.08190728724002838, + 0.019708750769495964, + 0.014425588771700859, + 0.021499885246157646, + 0.014899739995598793, + 0.02039947174489498, + 0.014787974767386913, + 0.010381192900240421, + 0.016477009281516075, + 0.00953293964266777, + 0.0067138671875, + 0.007055634167045355, + 0.006501362659037113, + 0.007465537171810865, + 0.008623902685940266, + 0.004428411368280649, + 0.009474135935306549, + 0.0056877294555306435, + 0.0034242868423461914, + 0.0036385157145559788, + 0.005215315148234367, + 0.006711628753691912, + 0.01991851069033146, + 0.018166087567806244, + 0.018663614988327026, + 0.014184897765517235, + 0.011445158161222935, + 0.008593437261879444, + 0.0163178201764822, + 0.011954891495406628, + 0.007310959510505199, + 0.007166683208197355, + 0.006035349797457457, + 0.007040726952254772, + 0.004545978270471096, + 0.00471400935202837, + 0.007431921549141407, + 0.008375565521419048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.01107108686119318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, static, panoramic shot transforms the scene into a rock painting, richly grained and mineral-hued. Across a vast wasteland of ochre and dark brown, two giant beetles, the size of armored vehicles, engage in a primal, ferocious struggle. Their obsidian shells gleam with a deep, bronze-metallic sheen, their surfaces crisscrossed with the characteristic rough texture and microscopic cracks of rock painting, as if composed of layers of the purest mineral pigment. One beetle thrusts forward, its six sturdy legs digging into the granular ground for grip, while the other tenses its body, stubbornly resisting the onslaught. The three exaggerated, curved horns atop their heads, like the blades of ancient warriors, are locked in a fierce, silent, and powerful struggle. Each clash unleashes a tremendous force, causing the air around their point of contact to distort and sending fine mineral dust soaring with the violent vibrations. Their movements are imbued with a primal sense of power. One attempts to pry and flip its opponent with its long horns, while the other lowers its body to stabilize its center of gravity. Every muscle contraction is clearly visible, as if one can feel the pulsation within through the rough shell. A dramatic, strong light from low above slants across the scene, carving sharp highlights and deep shadows into their metallic bodies, further intensifying the image's three-dimensionality and rugged texture, and creating an atmosphere of grand, tense, and epic confrontation.", + "seed": 7003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_13/index_7874_seed_7003_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": -0.013784529827535152, + "vlm_logprob": -0.013784529827535152, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.013784529827535152, + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.017796946689486504 + ], + "opposite_logprob_tokens": [ + -4.142796993255615 + ], + "logprob_reward": -0.017796946689486504, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017796946689486504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12484650313854218, + 0.123243547976017, + 0.07613734900951385, + 0.06282760947942734, + 0.03578706458210945, + 0.023693574592471123, + 0.3350466787815094, + 0.1331000179052353, + 0.06332163512706757, + 0.02199598029255867 + ], + "credit_map": [ + 0.024364711716771126, + 0.01192410197108984, + 0.01492689922451973, + 0.018948933109641075, + 0.01594528928399086, + 0.01525192242115736, + 0.01335997972637415, + 0.01012466847896576, + 0.01280394196510315, + 0.012127968482673168, + 0.034029554575681686, + 0.010776151902973652, + 0.008729837834835052, + 0.02351861260831356, + 0.011803414672613144, + 0.009454065933823586, + 0.0067092664539813995, + 0.011929525062441826, + 0.009478268213570118, + 0.01015572901815176, + 0.00813552737236023, + 0.010120588354766369, + 0.010281265713274479, + 0.009327179752290249, + 0.007910231128334999, + 0.005340577568858862, + 0.005897624883800745, + 0.009233568795025349, + 0.009271924383938313, + 0.00951257161796093, + 0.009528200142085552, + 0.006132906768471003, + 0.0055206166580319405, + 0.003637257032096386, + 0.0036347624845802784, + 0.004464744124561548, + 0.004587544593960047, + 0.0036133129615336657, + 0.006695011630654335, + 0.0036338167265057564, + 0.004204197321087122, + 0.003208301728591323, + 0.003483862616121769, + 0.002258139429613948, + 0.0025524948723614216, + 0.0015621724305674434, + 0.003741545369848609, + 0.0026828600093722343, + 0.09064794331789017, + 0.01159778144210577, + 0.13373616337776184, + 0.026232412084937096, + 0.006942525040358305, + 0.010380836203694344, + 0.017060797661542892, + 0.038448236882686615, + 0.023790685459971428, + 0.018251102417707443, + 0.022941749542951584, + 0.014159207232296467, + 0.011137084104120731, + 0.011763219721615314, + 0.015164023265242577, + 0.015892956405878067, + 0.01073604915291071, + 0.006809541955590248, + 0.010868766345083714, + 0.005173457320779562, + 0.005371034611016512, + 0.00825138296931982, + 0.008890503086149693, + 0.007220904808491468, + 0.0030530875083059072, + 0.0026629501953721046, + 0.0034328047186136246, + 0.0022252248600125313, + 0.002712452318519354, + 0.002994158770889044, + 0.0021283479873090982, + 0.0027869518380612135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.017796946689486504 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.011080282740294933 + ], + "opposite_logprob_tokens": [ + -4.511080265045166 + ], + "logprob_reward": -0.011080282740294933, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080282740294933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19143591821193695, + 0.07003402709960938, + 0.020838433876633644, + 0.038297127932310104, + 0.04009263589978218, + 0.037526536732912064, + 0.2650253474712372, + 0.12338808923959732, + 0.09078368544578552, + 0.12257827073335648 + ], + "credit_map": [ + 0.03784864395856857, + 0.016841402277350426, + 0.02148248814046383, + 0.022779244929552078, + 0.02348462864756584, + 0.03222667798399925, + 0.02140882983803749, + 0.015364006161689758, + 0.005304028745740652, + 0.006638043560087681, + 0.016832487657666206, + 0.005981323309242725, + 0.00546615943312645, + 0.018057378008961678, + 0.005796549841761589, + 0.005958053283393383, + 0.002211145590990782, + 0.003682468319311738, + 0.002853094832971692, + 0.0030829370953142643, + 0.002246338641270995, + 0.0023187119513750076, + 0.0021910518407821655, + 0.0022526877000927925, + 0.00448273541405797, + 0.0038929039146751165, + 0.0035990942269563675, + 0.004528227727860212, + 0.0058922129683196545, + 0.006609863601624966, + 0.005178602412343025, + 0.004113487433642149, + 0.005519493483006954, + 0.0048146238550543785, + 0.0049990639090538025, + 0.005189246963709593, + 0.0050482917577028275, + 0.0038305092602968216, + 0.00684441439807415, + 0.0038469939026981592, + 0.007987351156771183, + 0.0057807983830571175, + 0.0046620978973805904, + 0.0037595208268612623, + 0.003549329238012433, + 0.00219951244071126, + 0.005626684986054897, + 0.003961243201047182, + 0.07625692337751389, + 0.01047065295279026, + 0.10519015043973923, + 0.020313335582613945, + 0.005771035328507423, + 0.008028172887861729, + 0.013186367228627205, + 0.025808747857809067, + 0.017042851075530052, + 0.013011396862566471, + 0.016183065250515938, + 0.012644326314330101, + 0.012020069174468517, + 0.023662924766540527, + 0.013033351860940456, + 0.015790104866027832, + 0.014253600500524044, + 0.01074359379708767, + 0.014913266524672508, + 0.009650252759456635, + 0.010137252509593964, + 0.011132960207760334, + 0.010939065366983414, + 0.009013690985739231, + 0.019216930493712425, + 0.03677018731832504, + 0.011523373425006866, + 0.007277429569512606, + 0.007822341285645962, + 0.02178819291293621, + 0.009065577760338783, + 0.00911423284560442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.011080282740294933 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.014203879982233047 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014203879982233047, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203879982233047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20181018114089966, + 0.16607680916786194, + 0.03772333264350891, + 0.0934218093752861, + 0.10953264683485031, + 0.04607204720377922, + 0.15440154075622559, + 0.056114062666893005, + 0.08082010596990585, + 0.05402745306491852 + ], + "credit_map": [ + 0.03762989491224289, + 0.01822257600724697, + 0.02553412690758705, + 0.030224421992897987, + 0.019852375611662865, + 0.03216221183538437, + 0.01952720433473587, + 0.018657367676496506, + 0.012698213569819927, + 0.009998828172683716, + 0.04562588408589363, + 0.008793734945356846, + 0.011326213367283344, + 0.050702694803476334, + 0.015596194192767143, + 0.011335029266774654, + 0.004673660267144442, + 0.006455815862864256, + 0.004498396534472704, + 0.004873692989349365, + 0.004175265785306692, + 0.004842226859182119, + 0.004519905894994736, + 0.0036843670532107353, + 0.011055639944970608, + 0.00915583223104477, + 0.010465556755661964, + 0.012868582271039486, + 0.012924343347549438, + 0.01280966866761446, + 0.014897668734192848, + 0.009244514629244804, + 0.01719009131193161, + 0.011045468039810658, + 0.015358385629951954, + 0.010227369144558907, + 0.01592341437935829, + 0.009166426956653595, + 0.02042616717517376, + 0.010195320472121239, + 0.008505911566317081, + 0.005249902606010437, + 0.006364202126860619, + 0.0033534264657646418, + 0.004036706406623125, + 0.003591609187424183, + 0.007124967407435179, + 0.007845322601497173, + 0.039967454969882965, + 0.007518566679209471, + 0.05964939296245575, + 0.011991756036877632, + 0.003613715525716543, + 0.005474208388477564, + 0.007871377281844616, + 0.018315080553293228, + 0.009550418704748154, + 0.00837879627943039, + 0.009218123741447926, + 0.006885227747261524, + 0.004981183912605047, + 0.005018618889153004, + 0.005483833607286215, + 0.00659785745665431, + 0.015489276498556137, + 0.006674522068351507, + 0.011540447361767292, + 0.0057438453659415245, + 0.006908794865012169, + 0.007664504460990429, + 0.017138581722974777, + 0.009660138748586178, + 0.008720532059669495, + 0.00894103292375803, + 0.005626958329230547, + 0.003843551268801093, + 0.004565161652863026, + 0.008837563917040825, + 0.005173878278583288, + 0.008318775333464146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.014203879982233047 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.016064027324318886 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064027324318886, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064027324318886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09396974742412567, + 0.19514094293117523, + 0.051351822912693024, + 0.2819458246231079, + 0.017082231119275093, + 0.015471139922738075, + 0.06517557054758072, + 0.15736006200313568, + 0.07948222011327744, + 0.043020445853471756 + ], + "credit_map": [ + 0.013329753652215004, + 0.012551039457321167, + 0.011632127687335014, + 0.010840237140655518, + 0.010837829671800137, + 0.01842079870402813, + 0.008808589540421963, + 0.0075493729673326015, + 0.01636604219675064, + 0.009392048232257366, + 0.05744454264640808, + 0.0108717642724514, + 0.009910697117447853, + 0.06343013793230057, + 0.016093814745545387, + 0.011631899513304234, + 0.0038640100974589586, + 0.007603059057146311, + 0.004668462555855513, + 0.008506907150149345, + 0.00813103374093771, + 0.00604686513543129, + 0.006844775751233101, + 0.0056867110542953014, + 0.036579735577106476, + 0.033239372074604034, + 0.029234018176794052, + 0.03211945667862892, + 0.04088013246655464, + 0.0376729741692543, + 0.042004313319921494, + 0.030215803533792496, + 0.0024645815137773752, + 0.001624219585210085, + 0.0028750018682330847, + 0.0015235235914587975, + 0.0020222757011651993, + 0.0015819654799997807, + 0.00233095302246511, + 0.0026597094256430864, + 0.003035027766600251, + 0.002183722099289298, + 0.0019190633902326226, + 0.0013748530764132738, + 0.0017118772957473993, + 0.0012766337022185326, + 0.0022133986931294203, + 0.0017565644811838865, + 0.00939405057579279, + 0.004664302803575993, + 0.01389362197369337, + 0.00501715624704957, + 0.00342571665532887, + 0.0043297698721289635, + 0.00598288094624877, + 0.018468070775270462, + 0.022859418764710426, + 0.016694243997335434, + 0.01553376391530037, + 0.014882831834256649, + 0.02236245758831501, + 0.028018761426210403, + 0.017767904326319695, + 0.01924067735671997, + 0.008951031602919102, + 0.008540203794836998, + 0.009293542243540287, + 0.009799634106457233, + 0.008551228791475296, + 0.00957393366843462, + 0.016247903928160667, + 0.008524740114808083, + 0.00606036139652133, + 0.015233431942760944, + 0.003377649700269103, + 0.0021341159008443356, + 0.0030501431319862604, + 0.004227714147418737, + 0.0028611544985324144, + 0.006075876299291849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.016064027324318886 + }, + { + "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\": \"Are two giant beetles visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast wasteland of ochre and dark brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the beetles have obsidian shells with a bronze-metallic sheen and rough texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the beetles engaged in a struggle where one thrusts forward and the other resists?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dramatic and strong, coming from low above to create sharp highlights and deep 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": [ + -0.009777512401342392 + ], + "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": [ + 0.11829972267150879, + 0.061572201550006866, + 0.021292470395565033, + 0.039631374180316925, + 0.023194653913378716, + 0.04010184481739998, + 0.5137358903884888, + 0.09390074014663696, + 0.04531759023666382, + 0.04295341297984123 + ], + "credit_map": [ + 0.02241581119596958, + 0.010187794454395771, + 0.012503010220825672, + 0.014611170627176762, + 0.012968061491847038, + 0.020907891914248466, + 0.013147209770977497, + 0.01155878696590662, + 0.005748999770730734, + 0.005049790255725384, + 0.01436081063002348, + 0.004004903137683868, + 0.00332376966252923, + 0.015087653882801533, + 0.006916427984833717, + 0.007079850882291794, + 0.002563548507168889, + 0.0030384487472474575, + 0.0017929758178070188, + 0.0031617842614650726, + 0.002986295148730278, + 0.00256341858766973, + 0.002484883414581418, + 0.002701118355616927, + 0.00492433225736022, + 0.0034347539767622948, + 0.003975642845034599, + 0.005763702094554901, + 0.004827287048101425, + 0.004986753221601248, + 0.00694539537653327, + 0.004773510619997978, + 0.0035641121212393045, + 0.003104711649939418, + 0.005599357187747955, + 0.002358687575906515, + 0.002096563344821334, + 0.001615762710571289, + 0.0029712822288274765, + 0.0018841784913092852, + 0.008053726516664028, + 0.006201666779816151, + 0.005669252946972847, + 0.0036423138808459044, + 0.004179477691650391, + 0.0027123531326651573, + 0.005160342436283827, + 0.004482713993638754, + 0.1561969518661499, + 0.016932405531406403, + 0.2068946808576584, + 0.03969351947307587, + 0.009612222202122211, + 0.013841487467288971, + 0.025235677137970924, + 0.045329026877880096, + 0.015791265293955803, + 0.011327000334858894, + 0.011867332272231579, + 0.008264096453785896, + 0.010204331949353218, + 0.014904988929629326, + 0.009875676594674587, + 0.011666053906083107, + 0.008311167359352112, + 0.005150012206286192, + 0.0073430584743618965, + 0.004230191465467215, + 0.00485676946118474, + 0.005298859905451536, + 0.0052280607633292675, + 0.004899469669908285, + 0.005111369304358959, + 0.011397264897823334, + 0.0053702546283602715, + 0.0025741104036569595, + 0.002493578940629959, + 0.004588049370795488, + 0.0031632864847779274, + 0.00825550314038992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.009777512401342392 + } + ], + "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/0000167/group_14/index_tensor([441, 441])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_tensor([441, 441])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6bf678832945a657e99d66dbb611f0fc6ba5be21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_tensor([441, 441])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film's subsequent shots, executed in slow motion, track the propelling of a sleek, minimalist stainless steel ballpoint pen from the edge of a wooden workbench. The camera maintains tight focus on the pen as it gracefully arcs through the air in a perfectly elliptical trajectory. The studio is bathed in the warm golden light of late afternoon sunlight filtering through a large window, illuminating countless motes of dust dancing lazily in the sunlight. As the pen gently tumbles on its shaft, its polished cylindrical body catches the light, creating brilliant, fleeting specular highlights. The camera pans smoothly, keeping the pen perfectly centered and in sharp focus, while the background—a cluttered but blurred scene of carpentry tools, wood chips, and shelves on the far wall—dissolves completely into a soft, dreamy, cinematic bokeh composed of warm, circular blurs, creating an atmosphere of tranquility and visual poetry, focused entirely on the object's graceful, suspended movement.", + "seed": 23637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_441_seed_23637_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": -0.37532169427722695, + "vlm_logprob": -0.37532169427722695, + "vlm_yes_no_margin": 1.85, + "vlm_token_credit": -0.37532169427722695, + "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 main subject a sleek, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.23269112408161163 + ], + "opposite_logprob_tokens": [ + -1.6076911687850952 + ], + "logprob_reward": -0.23269112408161163, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23269112408161163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08096589148044586, + 0.09691682457923889, + 0.03138883411884308, + 0.2933999001979828, + 0.08685828000307083, + 0.06416161358356476, + 0.1360241025686264, + 0.11287301778793335, + 0.029716474935412407, + 0.06769507378339767 + ], + "credit_map": [ + 0.010080387815833092, + 0.010500520467758179, + 0.011885168962180614, + 0.009896238334476948, + 0.007867344655096531, + 0.010119994170963764, + 0.010195642709732056, + 0.010420591570436954, + 0.011368132196366787, + 0.01678384467959404, + 0.016447853296995163, + 0.015551120974123478, + 0.009572426788508892, + 0.009539877064526081, + 0.012651696801185608, + 0.005001874640583992, + 0.0033748408313840628, + 0.006193811539560556, + 0.007319177966564894, + 0.0029786787927150726, + 0.0032348353415727615, + 0.0025282183196395636, + 0.0027820784598588943, + 0.0029771931003779173, + 0.025739045813679695, + 0.03403538092970848, + 0.02911524288356304, + 0.03572999686002731, + 0.07566500455141068, + 0.03247706592082977, + 0.026055265218019485, + 0.03458291292190552, + 0.011833157390356064, + 0.010513389483094215, + 0.009235713630914688, + 0.009772846475243568, + 0.011417791247367859, + 0.012117516249418259, + 0.01012464053928852, + 0.011843220330774784, + 0.004206052981317043, + 0.007056685630232096, + 0.0071455128490924835, + 0.010657762177288532, + 0.00775525439530611, + 0.0076791043393313885, + 0.010498697869479656, + 0.009162535890936852, + 0.005245871376246214, + 0.040504638105630875, + 0.005370530299842358, + 0.008149721659719944, + 0.004554443061351776, + 0.0510702021420002, + 0.015056241303682327, + 0.006072455085813999, + 0.009887685999274254, + 0.016093183308839798, + 0.0179436132311821, + 0.013181721791625023, + 0.01399916596710682, + 0.014742003753781319, + 0.012641948647797108, + 0.014383695088326931, + 0.0032441976945847273, + 0.003653914900496602, + 0.0052582877688109875, + 0.0053725154139101505, + 0.00449368404224515, + 0.002525354502722621, + 0.0028490007389336824, + 0.002319521037861705, + 0.009202491492033005, + 0.006120937876403332, + 0.010052207857370377, + 0.007567861117422581, + 0.010321925394237041, + 0.008267284370958805, + 0.007717915344983339, + 0.008444446139037609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.23269112408161163 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538444519042969 + ], + "opposite_logprob_tokens": [ + -0.4288444519042969 + ], + "logprob_reward": -1.0538444519042969, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538444519042969, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04894881695508957, + 0.05057264864444733, + 0.032531678676605225, + 0.10128574073314667, + 0.026479989290237427, + 0.043075114488601685, + 0.5404013991355896, + 0.03583964332938194, + 0.038195230066776276, + 0.08266974985599518 + ], + "credit_map": [ + 0.006673440337181091, + 0.008267819881439209, + 0.005280510056763887, + 0.005898470524698496, + 0.00671788677573204, + 0.0047480640932917595, + 0.005638506729155779, + 0.005724118556827307, + 0.004513425286859274, + 0.007111827842891216, + 0.006011206191033125, + 0.005305962637066841, + 0.008092017844319344, + 0.00696531031280756, + 0.008993840776383877, + 0.003579056588932872, + 0.004634918179363012, + 0.009750647470355034, + 0.0038717722054570913, + 0.003070737002417445, + 0.0024565758649259806, + 0.003465858520939946, + 0.002116116927936673, + 0.0031650494784116745, + 0.011376443319022655, + 0.012010272592306137, + 0.01087704673409462, + 0.011990477330982685, + 0.023704297840595245, + 0.011489921249449253, + 0.007855251431465149, + 0.011982029303908348, + 0.0035099305678159, + 0.0036057562101632357, + 0.002360277809202671, + 0.0037922109477221966, + 0.0034638296347111464, + 0.003821834921836853, + 0.0031006699427962303, + 0.0028254790231585503, + 0.004074866883456707, + 0.0060381400398910046, + 0.004569138865917921, + 0.005053486675024033, + 0.0036333296447992325, + 0.004697203636169434, + 0.006465546786785126, + 0.008543401956558228, + 0.019533008337020874, + 0.16965577006340027, + 0.017072923481464386, + 0.025865405797958374, + 0.01715867966413498, + 0.2118135690689087, + 0.05820874124765396, + 0.021093349903821945, + 0.004208686761558056, + 0.004726117942482233, + 0.004227679688483477, + 0.005322632379829884, + 0.0032700716983526945, + 0.004194117616862059, + 0.005386023782193661, + 0.004504313692450523, + 0.00649953493848443, + 0.004276278428733349, + 0.005742063745856285, + 0.006306204479187727, + 0.003909966442734003, + 0.003538737539201975, + 0.003898350056260824, + 0.004024092108011246, + 0.009194616228342056, + 0.008267597295343876, + 0.013996484689414501, + 0.008491535671055317, + 0.0108803017064929, + 0.010586987249553204, + 0.009155194275081158, + 0.012097032740712166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -1.0538444519042969 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.0620780847966671 + ], + "opposite_logprob_tokens": [ + -2.8120779991149902 + ], + "logprob_reward": -0.0620780847966671, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620780847966671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01997780054807663, + 0.017075389623641968, + 0.0060822563245892525, + 0.01201821118593216, + 0.012481442652642727, + 0.009494928643107414, + 0.902306079864502, + 0.00772513123229146, + 0.004298121202737093, + 0.008540665730834007 + ], + "credit_map": [ + 0.0020867064595222473, + 0.0024939763825386763, + 0.0027470686472952366, + 0.003203899599611759, + 0.002665547886863351, + 0.0023827552795410156, + 0.0024168938398361206, + 0.001980951288715005, + 0.001980448141694069, + 0.0031522323843091726, + 0.0026504641864448786, + 0.0028419028967618942, + 0.001760898856446147, + 0.0017206184566020966, + 0.0018702404340729117, + 0.0010985842673107982, + 0.0006020030123181641, + 0.0017496730433776975, + 0.0015608671819791198, + 0.000601291423663497, + 0.0003600246272981167, + 0.000380681362003088, + 0.00034194288309663534, + 0.0004857726744376123, + 0.0010320364963263273, + 0.0014807257102802396, + 0.0014191584195941687, + 0.0015633952571079135, + 0.0015532674733549356, + 0.0014333840226754546, + 0.0014797478215768933, + 0.002056495053693652, + 0.0020808037370443344, + 0.0023750662803649902, + 0.001458896091207862, + 0.0014336459571495652, + 0.0012536559952422976, + 0.0010088789276778698, + 0.0015679645584896207, + 0.001302531803958118, + 0.0011849978473037481, + 0.0012686115223914385, + 0.0011738667963072658, + 0.001418032101355493, + 0.0006079226150177419, + 0.0012435066746547818, + 0.0010598921217024326, + 0.001538098556920886, + 0.03269819915294647, + 0.28223562240600586, + 0.02922622114419937, + 0.04621800035238266, + 0.028542686253786087, + 0.35139983892440796, + 0.09673689305782318, + 0.03524857759475708, + 0.000646038621198386, + 0.0011350527638569474, + 0.0014519266551360488, + 0.0010188458254560828, + 0.0009183157817460597, + 0.001019878312945366, + 0.0008017101208679378, + 0.0007333632674999535, + 0.0006956906290724874, + 0.000564836198464036, + 0.0006002389709465206, + 0.0007256371318362653, + 0.0004735477559734136, + 0.0004429227556101978, + 0.0004082661762367934, + 0.00038698146818205714, + 0.000977731542661786, + 0.0009160413173958659, + 0.0014652020763605833, + 0.0009004868916235864, + 0.0009560325415804982, + 0.0013051837449893355, + 0.0008412066381424665, + 0.0011787806870415807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0620780847966671 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.00464171078056097 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464171078056097, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464171078056097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16790837049484253, + 0.09848359227180481, + 0.05356154963374138, + 0.18824462592601776, + 0.04357583820819855, + 0.10617021471261978, + 0.20178629457950592, + 0.051166098564863205, + 0.0459216944873333, + 0.043181657791137695 + ], + "credit_map": [ + 0.023352380841970444, + 0.029060672968626022, + 0.015575041063129902, + 0.0158454030752182, + 0.026593796908855438, + 0.015977397561073303, + 0.020280536264181137, + 0.02122313529253006, + 0.00869221705943346, + 0.013957817107439041, + 0.011369557119905949, + 0.016736356541514397, + 0.017405414953827858, + 0.013300803489983082, + 0.01146245002746582, + 0.0055589755065739155, + 0.007448664866387844, + 0.006035824306309223, + 0.01604672707617283, + 0.006301309913396835, + 0.004157670773565769, + 0.003968331962823868, + 0.003915744833648205, + 0.005687273107469082, + 0.025709399953484535, + 0.02079351246356964, + 0.03384857997298241, + 0.01669486053287983, + 0.01744820550084114, + 0.01287342794239521, + 0.023752812296152115, + 0.03712384030222893, + 0.005402228329330683, + 0.004907519090920687, + 0.004238177090883255, + 0.005819183308631182, + 0.0052363029681146145, + 0.007133969571441412, + 0.005520948674529791, + 0.00531751150265336, + 0.008588986471295357, + 0.012732809409499168, + 0.011690521612763405, + 0.0174905676394701, + 0.006130018271505833, + 0.016118809580802917, + 0.010375690646469593, + 0.02304280921816826, + 0.007863313890993595, + 0.06236071512103081, + 0.006256969179958105, + 0.014032870531082153, + 0.006947922986000776, + 0.07481697201728821, + 0.0211926456540823, + 0.008314887061715126, + 0.004675686825066805, + 0.006950249429792166, + 0.0059696403332054615, + 0.008196292445063591, + 0.006184388883411884, + 0.00670333718881011, + 0.007194747216999531, + 0.005291758105158806, + 0.00728542311117053, + 0.004793981090188026, + 0.00540353637188673, + 0.007250020280480385, + 0.007161897607147694, + 0.004253532737493515, + 0.005164274480193853, + 0.004609032068401575, + 0.0048345415852963924, + 0.006164599675685167, + 0.00713737215846777, + 0.0048390342853963375, + 0.0034729766193777323, + 0.005376256536692381, + 0.003965210169553757, + 0.007391666062176228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.00464171078056097 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.523353099822998 + ], + "opposite_logprob_tokens": [ + -0.898353099822998 + ], + "logprob_reward": -0.523353099822998, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523353099822998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05248875170946121, + 0.028944173827767372, + 0.028265558183193207, + 0.03077409416437149, + 0.0322616845369339, + 0.04142275080084801, + 0.5879428386688232, + 0.10780379176139832, + 0.033640358597040176, + 0.05645599216222763 + ], + "credit_map": [ + 0.004921941086649895, + 0.0053915465250611305, + 0.012857062742114067, + 0.011750418692827225, + 0.0050182645209133625, + 0.004954686854034662, + 0.0038163145072758198, + 0.00377851165831089, + 0.003550961846485734, + 0.0034787659533321857, + 0.0041273655369877815, + 0.007383748423308134, + 0.002452449407428503, + 0.0024341163225471973, + 0.0035776798613369465, + 0.0019390865927562118, + 0.0032217693515121937, + 0.00529315322637558, + 0.00394365144893527, + 0.0036055801901966333, + 0.003178355284035206, + 0.0032090661115944386, + 0.002949156565591693, + 0.0028648243751376867, + 0.003007045015692711, + 0.0042625730857253075, + 0.003599011106416583, + 0.0030436834786087275, + 0.003925543278455734, + 0.0035037582274526358, + 0.004434229806065559, + 0.0049982438795268536, + 0.004247293341904879, + 0.003674166277050972, + 0.004018603824079037, + 0.0037549068219959736, + 0.004061346873641014, + 0.0034651735331863165, + 0.005203678738325834, + 0.003836512565612793, + 0.004301649983972311, + 0.004006619099527597, + 0.004294200800359249, + 0.007228517904877663, + 0.003449728712439537, + 0.006229056976735592, + 0.005911416374146938, + 0.006001555360853672, + 0.021599000319838524, + 0.18756942451000214, + 0.017112690955400467, + 0.027248114347457886, + 0.01866544596850872, + 0.22987572848796844, + 0.06315464526414871, + 0.022717786952853203, + 0.010664722882211208, + 0.015436246991157532, + 0.01975085772573948, + 0.01313008926808834, + 0.010867316275835037, + 0.014568351209163666, + 0.010755849070847034, + 0.012630347162485123, + 0.005210809409618378, + 0.0038076790515333414, + 0.005700922105461359, + 0.005581676959991455, + 0.0037171179428696632, + 0.003595979418605566, + 0.0031310580670833588, + 0.002895113080739975, + 0.0066257137805223465, + 0.006088390946388245, + 0.009820028208196163, + 0.0064040496945381165, + 0.005324062425643206, + 0.008209547027945518, + 0.005198088474571705, + 0.00878610834479332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.523353099822998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's subsequent shots, executed in slow motion, track the propelling of a sleek, minimalist stainless steel ballpoint pen from the edge of a wooden workbench. The camera maintains tight focus on the pen as it gracefully arcs through the air in a perfectly elliptical trajectory. The studio is bathed in the warm golden light of late afternoon sunlight filtering through a large window, illuminating countless motes of dust dancing lazily in the sunlight. As the pen gently tumbles on its shaft, its polished cylindrical body catches the light, creating brilliant, fleeting specular highlights. The camera pans smoothly, keeping the pen perfectly centered and in sharp focus, while the background—a cluttered but blurred scene of carpentry tools, wood chips, and shelves on the far wall—dissolves completely into a soft, dreamy, cinematic bokeh composed of warm, circular blurs, creating an atmosphere of tranquility and visual poetry, focused entirely on the object's graceful, suspended movement.", + "seed": 23638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_441_seed_23638_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": -0.3586193289607763, + "vlm_logprob": -0.3586193289607763, + "vlm_yes_no_margin": 1.900000011920929, + "vlm_token_credit": -0.3586193289607763, + "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 main subject a sleek, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.3977443277835846 + ], + "opposite_logprob_tokens": [ + -1.1477442979812622 + ], + "logprob_reward": -0.3977443277835846, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3977443277835846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044359851628541946, + 0.0381372906267643, + 0.07557497918605804, + 0.1402578502893448, + 0.24899153411388397, + 0.04610930755734444, + 0.02943417988717556, + 0.2787187099456787, + 0.06928139925003052, + 0.02913491614162922 + ], + "credit_map": [ + 0.0033666808158159256, + 0.004847404081374407, + 0.0059592342004179955, + 0.005693406332284212, + 0.004451198969036341, + 0.0034319839905947447, + 0.008610223419964314, + 0.007999720051884651, + 0.004811677150428295, + 0.005096616689115763, + 0.004592599347233772, + 0.004302648361772299, + 0.005727964919060469, + 0.003754860023036599, + 0.005502372048795223, + 0.004348548594862223, + 0.007477598264813423, + 0.009578815661370754, + 0.008794345892965794, + 0.00897687766700983, + 0.011275181546807289, + 0.008166962303221226, + 0.013775462284684181, + 0.007529732305556536, + 0.010288535617291927, + 0.010374883189797401, + 0.010200695134699345, + 0.013586128130555153, + 0.010949826799333096, + 0.03304053097963333, + 0.011148707941174507, + 0.0406685434281826, + 0.0219418965280056, + 0.02011648751795292, + 0.056747883558273315, + 0.040947992354631424, + 0.011545563116669655, + 0.040205344557762146, + 0.02661544643342495, + 0.03087090328335762, + 0.0032204410526901484, + 0.0056507885456085205, + 0.005184517707675695, + 0.009778792969882488, + 0.003901721676811576, + 0.005006392952054739, + 0.0076225255616009235, + 0.0057441238313913345, + 0.0031986471731215715, + 0.002543353708460927, + 0.004455056972801685, + 0.0037679895758628845, + 0.002729611936956644, + 0.005011105444282293, + 0.004022522829473019, + 0.0037058910820633173, + 0.03201885521411896, + 0.043063461780548096, + 0.03575829416513443, + 0.03309199586510658, + 0.030770935118198395, + 0.027093876153230667, + 0.035348180681467056, + 0.04157310724258423, + 0.007115599699318409, + 0.011493555270135403, + 0.013173965737223625, + 0.008272454142570496, + 0.00678679533302784, + 0.006821704097092152, + 0.008612916804850101, + 0.007004404440522194, + 0.002196803456172347, + 0.00311434012837708, + 0.002510210033506155, + 0.0031140365172177553, + 0.004620616789907217, + 0.005110795143991709, + 0.005691912490874529, + 0.002776197623461485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.3977443277835846 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9742588996887207 + ], + "opposite_logprob_tokens": [ + -0.4742588996887207 + ], + "logprob_reward": -0.9742588996887207, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742588996887207, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05278947576880455, + 0.0989493653178215, + 0.06014648824930191, + 0.33362695574760437, + 0.05480445548892021, + 0.03924769163131714, + 0.05359656363725662, + 0.13124093413352966, + 0.03908924013376236, + 0.13650880753993988 + ], + "credit_map": [ + 0.005106330383569002, + 0.0062110829167068005, + 0.0035517970100045204, + 0.006102017126977444, + 0.007948474027216434, + 0.008544965647161007, + 0.00708351843059063, + 0.008241293020546436, + 0.008170189335942268, + 0.010256479494273663, + 0.011229261755943298, + 0.011513794772326946, + 0.012652961537241936, + 0.013921507634222507, + 0.014747781679034233, + 0.016457388177514076, + 0.009624764323234558, + 0.008097389712929726, + 0.007377466652542353, + 0.007964643649756908, + 0.0048716128803789616, + 0.006719536148011684, + 0.007852787151932716, + 0.007638289127498865, + 0.019958192482590675, + 0.021573299542069435, + 0.05433846265077591, + 0.03025149554014206, + 0.02376164309680462, + 0.058426257222890854, + 0.02406921796500683, + 0.1012483760714531, + 0.004617529921233654, + 0.005060269962996244, + 0.009197810664772987, + 0.01056510116904974, + 0.0022840106394141912, + 0.00856058206409216, + 0.005704841110855341, + 0.008814310654997826, + 0.005128693301230669, + 0.00483580119907856, + 0.006193042732775211, + 0.005386176984757185, + 0.003015566850081086, + 0.004101475700736046, + 0.005356399342417717, + 0.0052305362187325954, + 0.0040272679179906845, + 0.005168300122022629, + 0.007501327432692051, + 0.008352791890501976, + 0.005182311870157719, + 0.004834527149796486, + 0.011522158980369568, + 0.007007880136370659, + 0.01199694350361824, + 0.016984859481453896, + 0.016258984804153442, + 0.016905121505260468, + 0.016062617301940918, + 0.014249355532228947, + 0.017703521996736526, + 0.02107953280210495, + 0.003714196151122451, + 0.007720037829130888, + 0.0037773791700601578, + 0.005065056029707193, + 0.004750307649374008, + 0.0036164522171020508, + 0.006683390121906996, + 0.0037624246906489134, + 0.0058992658741772175, + 0.013937709853053093, + 0.010702547617256641, + 0.016138646751642227, + 0.017756221815943718, + 0.025311019271612167, + 0.032161880284547806, + 0.014601511880755424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.9742588996887207 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.02635873481631279 + ], + "opposite_logprob_tokens": [ + -3.6513588428497314 + ], + "logprob_reward": -0.02635873481631279, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635873481631279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06257965415716171, + 0.11341819167137146, + 0.20687705278396606, + 0.12017960101366043, + 0.11545751243829727, + 0.08327082544565201, + 0.05787668377161026, + 0.0838453397154808, + 0.02677365578711033, + 0.1297215223312378 + ], + "credit_map": [ + 0.004261455032974482, + 0.005702688358724117, + 0.005664863158017397, + 0.007252839393913746, + 0.007266182918101549, + 0.006524547003209591, + 0.011751091107726097, + 0.014155984856188297, + 0.006770918145775795, + 0.009931439533829689, + 0.011639903299510479, + 0.013860046863555908, + 0.012049706652760506, + 0.012278764508664608, + 0.03426226228475571, + 0.012625138275325298, + 0.021346481516957283, + 0.022358188405632973, + 0.02962176315486431, + 0.0323706790804863, + 0.012119336985051632, + 0.033802662044763565, + 0.01489211805164814, + 0.04036582633852959, + 0.009013241156935692, + 0.008853710256516933, + 0.028719697147607803, + 0.010967142879962921, + 0.007425718940794468, + 0.014403307810425758, + 0.0064698136411607265, + 0.03432696312665939, + 0.00953578855842352, + 0.00920993834733963, + 0.031215166673064232, + 0.02063147723674774, + 0.004384809173643589, + 0.016651973128318787, + 0.012005334720015526, + 0.011823028326034546, + 0.01280790101736784, + 0.008549184538424015, + 0.0141343604773283, + 0.01220185961574316, + 0.0067977746948599815, + 0.007461395114660263, + 0.011521969921886921, + 0.009796378202736378, + 0.0047306171618402, + 0.004992655478417873, + 0.01370866596698761, + 0.007426680531352758, + 0.0038283958565443754, + 0.004684589337557554, + 0.00797924306243658, + 0.010525836609303951, + 0.008145853877067566, + 0.014048482291400433, + 0.00844537653028965, + 0.011550400406122208, + 0.008772814646363258, + 0.0066412389278411865, + 0.011782627552747726, + 0.014458546414971352, + 0.003104256000369787, + 0.0034205822739750147, + 0.0034485969226807356, + 0.0044237589463591576, + 0.0020898361690342426, + 0.0023231254890561104, + 0.003791647730395198, + 0.00417185015976429, + 0.015388659201562405, + 0.01552086602896452, + 0.014983627945184708, + 0.017475297674536705, + 0.011378052644431591, + 0.014177361503243446, + 0.018055692315101624, + 0.022741980850696564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.02635873481631279 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07846343517303467, + 0.13690167665481567, + 0.09373143315315247, + 0.13859069347381592, + 0.06790661811828613, + 0.17924301326274872, + 0.047797635197639465, + 0.07623779773712158, + 0.06875869631767273, + 0.11236898601055145 + ], + "credit_map": [ + 0.0065515777096152306, + 0.008262517862021923, + 0.005511139985173941, + 0.007518939208239317, + 0.008867470547556877, + 0.009266357868909836, + 0.015522455796599388, + 0.016962973400950432, + 0.01175143476575613, + 0.010354946367442608, + 0.012849731370806694, + 0.021352292969822884, + 0.008659117855131626, + 0.006242712493985891, + 0.04902234673500061, + 0.016669075936079025, + 0.017484966665506363, + 0.012518162839114666, + 0.012942405417561531, + 0.01203639805316925, + 0.005697406828403473, + 0.011044749990105629, + 0.009534820914268494, + 0.012472512200474739, + 0.0073253377340734005, + 0.00692096259444952, + 0.051703039556741714, + 0.013555550947785378, + 0.0057226684875786304, + 0.011633433401584625, + 0.00920087844133377, + 0.03252880647778511, + 0.007755619008094072, + 0.0060172607190907, + 0.019278090447187424, + 0.009273534640669823, + 0.0033322798553854227, + 0.006234019994735718, + 0.006285384297370911, + 0.009730425663292408, + 0.013945612125098705, + 0.021764587610960007, + 0.01816377602517605, + 0.03353900834918022, + 0.014267927035689354, + 0.0192855317145586, + 0.02808382920920849, + 0.030192725360393524, + 0.004040232393890619, + 0.003222100902348757, + 0.014379385858774185, + 0.004431923385709524, + 0.002648334251716733, + 0.004213864915072918, + 0.0033425665460526943, + 0.01151922158896923, + 0.008811508305370808, + 0.01415216363966465, + 0.007569990120828152, + 0.009011129848659039, + 0.009786971844732761, + 0.00665231142193079, + 0.010131049901247025, + 0.010122663341462612, + 0.005335817113518715, + 0.007412251550704241, + 0.007956882007420063, + 0.010179017670452595, + 0.00664835050702095, + 0.00684670265763998, + 0.011251628398895264, + 0.013128041289746761, + 0.009264133870601654, + 0.011724283918738365, + 0.013381561264395714, + 0.013501740992069244, + 0.012758363038301468, + 0.014437710866332054, + 0.01999019831418991, + 0.017310990020632744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.007632853463292122 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.38710182905197144 + ], + "opposite_logprob_tokens": [ + -1.1371018886566162 + ], + "logprob_reward": -0.38710182905197144, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38710182905197144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05872849375009537, + 0.07905000448226929, + 0.06729636341333389, + 0.07434050738811493, + 0.34432464838027954, + 0.09408017247915268, + 0.04766607657074928, + 0.12086658179759979, + 0.02597486786544323, + 0.08767230063676834 + ], + "credit_map": [ + 0.004873158875852823, + 0.0076847979798913, + 0.007120312191545963, + 0.008423199877142906, + 0.007817397825419903, + 0.006483201868832111, + 0.00919112004339695, + 0.007135310210287571, + 0.006153787951916456, + 0.011372980661690235, + 0.008443101309239864, + 0.009136113338172436, + 0.008513441309332848, + 0.00897847581654787, + 0.014384552836418152, + 0.012067550793290138, + 0.00788550078868866, + 0.006819864269345999, + 0.008796818554401398, + 0.008460985496640205, + 0.005391414742916822, + 0.012283208779990673, + 0.006753070279955864, + 0.010905499570071697, + 0.0028321691788733006, + 0.00450212461873889, + 0.021451959386467934, + 0.004653331823647022, + 0.0032774643041193485, + 0.008629268035292625, + 0.005373741965740919, + 0.023620447143912315, + 0.031762056052684784, + 0.03066641092300415, + 0.06961030513048172, + 0.060046128928661346, + 0.015410389751195908, + 0.053296420723199844, + 0.03510664775967598, + 0.04842628166079521, + 0.015139986760914326, + 0.010924965143203735, + 0.010990513488650322, + 0.011339491233229637, + 0.007032306399196386, + 0.009974033571779728, + 0.013385537080466747, + 0.015293334610760212, + 0.004160214215517044, + 0.005078084301203489, + 0.012550869956612587, + 0.005774470046162605, + 0.004177495837211609, + 0.0063001313246786594, + 0.005159769207239151, + 0.004465041216462851, + 0.012246382422745228, + 0.023466961458325386, + 0.012125429697334766, + 0.014430086128413677, + 0.015042118728160858, + 0.011176514439284801, + 0.016318682581186295, + 0.01606040820479393, + 0.003119874279946089, + 0.004097281023859978, + 0.002632653806358576, + 0.003176341997459531, + 0.0025310534983873367, + 0.003131090197712183, + 0.00424391683191061, + 0.003042655298486352, + 0.005595541559159756, + 0.007853299379348755, + 0.00954858586192131, + 0.012766741216182709, + 0.01047548372298479, + 0.01346947718411684, + 0.014412164688110352, + 0.013551007024943829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.38710182905197144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's subsequent shots, executed in slow motion, track the propelling of a sleek, minimalist stainless steel ballpoint pen from the edge of a wooden workbench. The camera maintains tight focus on the pen as it gracefully arcs through the air in a perfectly elliptical trajectory. The studio is bathed in the warm golden light of late afternoon sunlight filtering through a large window, illuminating countless motes of dust dancing lazily in the sunlight. As the pen gently tumbles on its shaft, its polished cylindrical body catches the light, creating brilliant, fleeting specular highlights. The camera pans smoothly, keeping the pen perfectly centered and in sharp focus, while the background—a cluttered but blurred scene of carpentry tools, wood chips, and shelves on the far wall—dissolves completely into a soft, dreamy, cinematic bokeh composed of warm, circular blurs, creating an atmosphere of tranquility and visual poetry, focused entirely on the object's graceful, suspended movement.", + "seed": 23639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_441_seed_23639_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": -0.37811433165334163, + "vlm_logprob": -0.37811433165334163, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.37811433165334163, + "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 main subject a sleek, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.2597065567970276 + ], + "opposite_logprob_tokens": [ + -1.5097064971923828 + ], + "logprob_reward": -0.2597065567970276, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2597065567970276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1129734143614769, + 0.0580277293920517, + 0.1572750061750412, + 0.11959165334701538, + 0.08617772161960602, + 0.04047810286283493, + 0.09111911803483963, + 0.19424612820148468, + 0.08429687470197678, + 0.055814266204833984 + ], + "credit_map": [ + 0.013203189708292484, + 0.014003332704305649, + 0.020303497090935707, + 0.02170657180249691, + 0.011785323731601238, + 0.007340865675359964, + 0.012555358931422234, + 0.012075266800820827, + 0.00465375417843461, + 0.0036302690859884024, + 0.00916679110378027, + 0.0171816349029541, + 0.005588881205767393, + 0.005624309182167053, + 0.005432135425508022, + 0.006749948486685753, + 0.016070665791630745, + 0.031745538115501404, + 0.025100113824009895, + 0.009173497557640076, + 0.010023882612586021, + 0.012210167944431305, + 0.01644444838166237, + 0.03650669381022453, + 0.010208393447101116, + 0.010769522748887539, + 0.013316317461431026, + 0.012072359211742878, + 0.01902494765818119, + 0.016734912991523743, + 0.016434431076049805, + 0.02103075385093689, + 0.006921790540218353, + 0.010842494666576385, + 0.020729834213852882, + 0.010216576047241688, + 0.007489031180739403, + 0.016337916254997253, + 0.008420311845839024, + 0.005219754762947559, + 0.003639700822532177, + 0.005537761375308037, + 0.005825007800012827, + 0.006945648696273565, + 0.0053749242797493935, + 0.005732913035899401, + 0.0037945692893117666, + 0.0036275768652558327, + 0.024727165699005127, + 0.020961923524737358, + 0.009156699292361736, + 0.005645486526191235, + 0.0051848734728991985, + 0.010649490170180798, + 0.0057618506252765656, + 0.009031624533236027, + 0.02133706584572792, + 0.017627110704779625, + 0.025335876271128654, + 0.07267209142446518, + 0.01473158411681652, + 0.016183199360966682, + 0.013131307438015938, + 0.013227880001068115, + 0.006905368994921446, + 0.015477119944989681, + 0.015504375100135803, + 0.015059879049658775, + 0.007498737890273333, + 0.006099347956478596, + 0.009794601239264011, + 0.007957437075674534, + 0.005227074027061462, + 0.003305463120341301, + 0.014986722730100155, + 0.007290343753993511, + 0.0036243796348571777, + 0.005377973895519972, + 0.004100263584405184, + 0.011902041733264923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.2597065567970276 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235795259475708 + ], + "opposite_logprob_tokens": [ + -0.3485795259475708 + ], + "logprob_reward": -1.2235795259475708, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2235795259475708, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13784156739711761, + 0.08401593565940857, + 0.11925209313631058, + 0.22324182093143463, + 0.08594009280204773, + 0.0490553043782711, + 0.037937816232442856, + 0.1780179888010025, + 0.04152819514274597, + 0.04316914454102516 + ], + "credit_map": [ + 0.01559332013130188, + 0.01800360344350338, + 0.029292460530996323, + 0.02530166134238243, + 0.009815443307161331, + 0.010035241954028606, + 0.016814831644296646, + 0.012985005974769592, + 0.0094791604205966, + 0.007793628145009279, + 0.010969740338623524, + 0.015116970986127853, + 0.007629447150975466, + 0.010994893498718739, + 0.008867615833878517, + 0.013164477422833443, + 0.010407713241875172, + 0.01936715841293335, + 0.018258821219205856, + 0.013663413003087044, + 0.008840027265250683, + 0.010910052806138992, + 0.011371957138180733, + 0.026432963088154793, + 0.019398437812924385, + 0.024266811087727547, + 0.018784377723932266, + 0.021745041012763977, + 0.04810088500380516, + 0.02852543070912361, + 0.02925729751586914, + 0.03316355496644974, + 0.01081103179603815, + 0.007781027350574732, + 0.023205237463116646, + 0.0072929998859763145, + 0.007601313292980194, + 0.014347169548273087, + 0.008261967450380325, + 0.006639339495450258, + 0.004212291911244392, + 0.004910469055175781, + 0.01428732369095087, + 0.0059889270924031734, + 0.004008854739367962, + 0.007754562888294458, + 0.0024043286684900522, + 0.005488546099513769, + 0.00824823696166277, + 0.003981515299528837, + 0.005979439243674278, + 0.0044700466096401215, + 0.003557146992534399, + 0.0034509829711169004, + 0.003009159117937088, + 0.005241288337856531, + 0.02352716214954853, + 0.015101262368261814, + 0.025298209860920906, + 0.0593753382563591, + 0.009151160717010498, + 0.014204435981810093, + 0.007919250056147575, + 0.023441165685653687, + 0.00670254323631525, + 0.006061510648578405, + 0.013136337511241436, + 0.002960176207125187, + 0.0023003697860985994, + 0.003188328817486763, + 0.003330222563818097, + 0.0038487038109451532, + 0.005361245013773441, + 0.004014717414975166, + 0.011715635657310486, + 0.005360775627195835, + 0.003287134924903512, + 0.003993022255599499, + 0.005682059563696384, + 0.003754554083570838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -1.2235795259475708 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.05499187111854553 + ], + "opposite_logprob_tokens": [ + -2.9299919605255127 + ], + "logprob_reward": -0.05499187111854553, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499187111854553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08338458836078644, + 0.16636690497398376, + 0.1398966908454895, + 0.028844013810157776, + 0.28623130917549133, + 0.02777809090912342, + 0.06540144234895706, + 0.11741221696138382, + 0.01991351693868637, + 0.06477116793394089 + ], + "credit_map": [ + 0.007989242672920227, + 0.012872805818915367, + 0.009520171210169792, + 0.013913520611822605, + 0.008255194872617722, + 0.008076893165707588, + 0.012895943596959114, + 0.009860830381512642, + 0.023620566353201866, + 0.011849571950733662, + 0.021663812920451164, + 0.0310794897377491, + 0.012074330821633339, + 0.026551472023129463, + 0.014552968554198742, + 0.02497471123933792, + 0.013990355655550957, + 0.02649628557264805, + 0.023937927559018135, + 0.007090720348060131, + 0.008287286385893822, + 0.01110098697245121, + 0.014841906726360321, + 0.034151241183280945, + 0.003539704019203782, + 0.00285701104439795, + 0.00442355964332819, + 0.003051245352253318, + 0.004129706881940365, + 0.004152110777795315, + 0.003195052733644843, + 0.0034956266172230244, + 0.038097936660051346, + 0.016764404252171516, + 0.11378591507673264, + 0.013796312734484673, + 0.01046756375581026, + 0.05451175943017006, + 0.01575826108455658, + 0.02304919622838497, + 0.003375211963430047, + 0.004352778196334839, + 0.004097541328519583, + 0.0031348534394055605, + 0.00342645519413054, + 0.004026809707283974, + 0.0023081903345882893, + 0.0030562535393983126, + 0.01839863322675228, + 0.009019773453474045, + 0.011206666938960552, + 0.004206942860037088, + 0.0032900753431022167, + 0.008531641215085983, + 0.0035851732827723026, + 0.007162545807659626, + 0.013950151391327381, + 0.011117869056761265, + 0.0199026670306921, + 0.021998371928930283, + 0.008218923583626747, + 0.00944562442600727, + 0.007624630816280842, + 0.025153998285531998, + 0.0029639971908181906, + 0.0022314207162708044, + 0.003980292472988367, + 0.002739265561103821, + 0.001720787608064711, + 0.0016489375848323107, + 0.0021260729990899563, + 0.002502744784578681, + 0.004701824393123388, + 0.003423479851335287, + 0.020593341439962387, + 0.004039211664348841, + 0.0021686742547899485, + 0.002708599902689457, + 0.005527000408619642, + 0.021609047427773476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.05499187111854553 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.003617767943069339 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617767943069339, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617767943069339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10297391563653946, + 0.1298678070306778, + 0.2095007449388504, + 0.07791445404291153, + 0.03562207147479057, + 0.06439810991287231, + 0.09221521764993668, + 0.09788495302200317, + 0.03313582390546799, + 0.15648691356182098 + ], + "credit_map": [ + 0.010090711526572704, + 0.008957944810390472, + 0.018651003018021584, + 0.023554200306534767, + 0.009955385699868202, + 0.00803396850824356, + 0.012815142050385475, + 0.010915554128587246, + 0.012302841991186142, + 0.010297950357198715, + 0.019464410841464996, + 0.034421004354953766, + 0.00867293681949377, + 0.012160598300397396, + 0.013685861602425575, + 0.018862193450331688, + 0.02242784947156906, + 0.04621162265539169, + 0.020910335704684258, + 0.0154198557138443, + 0.017522528767585754, + 0.016889147460460663, + 0.019572479650378227, + 0.050546929240226746, + 0.006520642898976803, + 0.009012916125357151, + 0.007267871871590614, + 0.007483636494725943, + 0.015356602147221565, + 0.010119037702679634, + 0.010665470734238625, + 0.011488271877169609, + 0.0028486615046858788, + 0.005343947093933821, + 0.005825365427881479, + 0.004231407307088375, + 0.0035789681132882833, + 0.006663612090051174, + 0.004146842751652002, + 0.002983264159411192, + 0.005009162705391645, + 0.0052223848178982735, + 0.01784323900938034, + 0.008168659172952175, + 0.003951192367821932, + 0.012469162233173847, + 0.0025132265873253345, + 0.009221085347235203, + 0.02362963743507862, + 0.014087247662246227, + 0.014694365672767162, + 0.004500290844589472, + 0.0038860461208969355, + 0.012846168130636215, + 0.005625160411000252, + 0.012946298345923424, + 0.012481998652219772, + 0.008485134690999985, + 0.013857782818377018, + 0.022003047168254852, + 0.006837924011051655, + 0.006524472497403622, + 0.007171960081905127, + 0.020522629842162132, + 0.006980828940868378, + 0.004434895236045122, + 0.007745805196464062, + 0.0032835376914590597, + 0.00180432153865695, + 0.0023801533970981836, + 0.0023590209893882275, + 0.004147259518504143, + 0.013688911683857441, + 0.009833728894591331, + 0.027728775516152382, + 0.011642956174910069, + 0.005764458328485489, + 0.006660273764282465, + 0.025499118492007256, + 0.05566868931055069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.003617767943069339 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.348675936460495 + ], + "opposite_logprob_tokens": [ + -1.2236759662628174 + ], + "logprob_reward": -0.348675936460495, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.348675936460495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13115614652633667, + 0.07797010242938995, + 0.05352333188056946, + 0.07357028871774673, + 0.482381135225296, + 0.045196689665317535, + 0.027950741350650787, + 0.036425717175006866, + 0.028065389022231102, + 0.04376047104597092 + ], + "credit_map": [ + 0.00945995096117258, + 0.009664751589298248, + 0.04061180353164673, + 0.03387552127242088, + 0.012175522744655609, + 0.007818170823156834, + 0.008133570663630962, + 0.00941684190183878, + 0.006451979745179415, + 0.008027083240449429, + 0.009249857626855373, + 0.018251312896609306, + 0.008058504201471806, + 0.010097137652337551, + 0.010185142047703266, + 0.007649085484445095, + 0.005751211661845446, + 0.011584549210965633, + 0.005190495867282152, + 0.00331531697884202, + 0.004083288833498955, + 0.005314873997122049, + 0.00560541870072484, + 0.012678179889917374, + 0.005607970990240574, + 0.007414080668240786, + 0.008156750351190567, + 0.00936471950262785, + 0.01362957339733839, + 0.008889783173799515, + 0.010197622701525688, + 0.010309789329767227, + 0.06644889712333679, + 0.027833392843604088, + 0.19133830070495605, + 0.02319369651377201, + 0.0148869464173913, + 0.0916597992181778, + 0.029541965574026108, + 0.037478141486644745, + 0.005336180329322815, + 0.006184819154441357, + 0.008211405016481876, + 0.006814741063863039, + 0.004231420811265707, + 0.006565297022461891, + 0.0033146508503705263, + 0.004538177512586117, + 0.004544899333268404, + 0.003357985056936741, + 0.0030786958523094654, + 0.0036521952133625746, + 0.003450685413554311, + 0.004180028568953276, + 0.0024065845645964146, + 0.003279668977484107, + 0.004267936572432518, + 0.0037992901634424925, + 0.005604866426438093, + 0.004987670574337244, + 0.0035023600794374943, + 0.002976455260068178, + 0.0034084173385053873, + 0.007878720760345459, + 0.003305914346128702, + 0.00413775397464633, + 0.007769917603582144, + 0.004043507389724255, + 0.00240290816873312, + 0.0018437239341437817, + 0.002629325957968831, + 0.0019323363667353988, + 0.005461325868964195, + 0.004510380327701569, + 0.008687649853527546, + 0.006499918643385172, + 0.0037211969029158354, + 0.004898774437606335, + 0.004948562942445278, + 0.005032661836594343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.348675936460495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film's subsequent shots, executed in slow motion, track the propelling of a sleek, minimalist stainless steel ballpoint pen from the edge of a wooden workbench. The camera maintains tight focus on the pen as it gracefully arcs through the air in a perfectly elliptical trajectory. The studio is bathed in the warm golden light of late afternoon sunlight filtering through a large window, illuminating countless motes of dust dancing lazily in the sunlight. As the pen gently tumbles on its shaft, its polished cylindrical body catches the light, creating brilliant, fleeting specular highlights. The camera pans smoothly, keeping the pen perfectly centered and in sharp focus, while the background—a cluttered but blurred scene of carpentry tools, wood chips, and shelves on the far wall—dissolves completely into a soft, dreamy, cinematic bokeh composed of warm, circular blurs, creating an atmosphere of tranquility and visual poetry, focused entirely on the object's graceful, suspended movement.", + "seed": 23640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_14/index_441_seed_23640_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": -0.30357352253049613, + "vlm_logprob": -0.30357352253049613, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.30357352253049613, + "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 main subject a sleek, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.05406103655695915 + ], + "opposite_logprob_tokens": [ + -3.304060935974121 + ], + "logprob_reward": -0.05406103655695915, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05406103655695915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05273091420531273, + 0.019060034304857254, + 0.03730000555515289, + 0.05925516039133072, + 0.05846662074327469, + 0.18724113702774048, + 0.05124581232666969, + 0.0622236430644989, + 0.37947240471839905, + 0.09300423413515091 + ], + "credit_map": [ + 0.007147868629544973, + 0.009235545061528683, + 0.004180074669420719, + 0.004615379497408867, + 0.008360411040484905, + 0.005207050126045942, + 0.008582993410527706, + 0.0054015954956412315, + 0.003473597811535001, + 0.003089484991505742, + 0.0017448693979531527, + 0.0020912904292345047, + 0.0030702203512191772, + 0.0031867127399891615, + 0.0011895527131855488, + 0.001214308082126081, + 0.0031213555485010147, + 0.004962103441357613, + 0.011964871548116207, + 0.002746597398072481, + 0.004049477633088827, + 0.003755462821573019, + 0.0021754896733909845, + 0.004524653311818838, + 0.021209346130490303, + 0.02194163203239441, + 0.001679019071161747, + 0.0014667725190520287, + 0.004660769831389189, + 0.0034872873220592737, + 0.003214340889826417, + 0.001595995738171041, + 0.005732299294322729, + 0.004794988315552473, + 0.012877089902758598, + 0.0034067295491695404, + 0.010006580501794815, + 0.005588939879089594, + 0.007172885350883007, + 0.008887110278010368, + 0.04933733865618706, + 0.015324029140174389, + 0.016467519104480743, + 0.00900548417121172, + 0.006568167358636856, + 0.012179886922240257, + 0.013490217737853527, + 0.0648685023188591, + 0.004889606963843107, + 0.007278812117874622, + 0.004652951844036579, + 0.005311867687851191, + 0.0068876175209879875, + 0.009241233579814434, + 0.006305146496742964, + 0.006678578443825245, + 0.008361920714378357, + 0.010573839768767357, + 0.007075396832078695, + 0.005933168809860945, + 0.00870555266737938, + 0.007829942740499973, + 0.008790559135377407, + 0.0049532633274793625, + 0.1269555687904358, + 0.14277321100234985, + 0.014731903560459614, + 0.009414572268724442, + 0.03381914272904396, + 0.027750149369239807, + 0.013877127319574356, + 0.010150747373700142, + 0.007911915890872478, + 0.015172875486314297, + 0.007054392248392105, + 0.00847335159778595, + 0.02525296062231064, + 0.011134716682136059, + 0.010514124296605587, + 0.007489894982427359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.05406103655695915 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137001633644104 + ], + "opposite_logprob_tokens": [ + -0.3870016038417816 + ], + "logprob_reward": -1.137001633644104, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137001633644104, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11910196393728256, + 0.03575332462787628, + 0.0637170597910881, + 0.19297510385513306, + 0.09702571481466293, + 0.13416987657546997, + 0.048946406692266464, + 0.05491267889738083, + 0.2087770253419876, + 0.04462076723575592 + ], + "credit_map": [ + 0.016595831140875816, + 0.0196668840944767, + 0.007595086470246315, + 0.008076293393969536, + 0.030138695612549782, + 0.01391400396823883, + 0.013264643959701061, + 0.009850536473095417, + 0.008526407182216644, + 0.008309697732329369, + 0.0025406833738088608, + 0.0054436917416751385, + 0.0043135941959917545, + 0.003382332855835557, + 0.0018011837964877486, + 0.0014357361942529678, + 0.008042233996093273, + 0.010165635496377945, + 0.020091939717531204, + 0.0072594997473061085, + 0.00520292529836297, + 0.005688286852091551, + 0.002862224355340004, + 0.004404312465339899, + 0.07235277444124222, + 0.07320566475391388, + 0.006041407119482756, + 0.006080535240471363, + 0.0075212763622403145, + 0.0077580553479492664, + 0.009616497904062271, + 0.010398896411061287, + 0.007853466086089611, + 0.0060456725768744946, + 0.032619595527648926, + 0.005512826144695282, + 0.010130973532795906, + 0.0034169836435467005, + 0.009493055753409863, + 0.021953146904706955, + 0.0323508195579052, + 0.010900374501943588, + 0.01341077871620655, + 0.007287891581654549, + 0.00627540610730648, + 0.008062970824539661, + 0.01134559977799654, + 0.0445360504090786, + 0.004778564907610416, + 0.0052335867658257484, + 0.004654021468013525, + 0.0054657962173223495, + 0.005436231382191181, + 0.01075708493590355, + 0.00558109488338232, + 0.007040031719952822, + 0.008639990352094173, + 0.00790735986083746, + 0.0063276272267103195, + 0.005124474409967661, + 0.008249478414654732, + 0.009107881225645542, + 0.004961568862199783, + 0.004594305530190468, + 0.06401211023330688, + 0.07729839533567429, + 0.01163493748754263, + 0.005592270288616419, + 0.018514981493353844, + 0.01614583656191826, + 0.007792947348207235, + 0.007785562891513109, + 0.0044444529339671135, + 0.007727091666311026, + 0.004318711813539267, + 0.0077866194769740105, + 0.006950927898287773, + 0.004954707343131304, + 0.002986367791891098, + 0.0054518962278962135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -1.137001633644104 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.07004504650831223 + ], + "opposite_logprob_tokens": [ + -2.695044994354248 + ], + "logprob_reward": -0.07004504650831223, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004504650831223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14930060505867004, + 0.12410883605480194, + 0.05360742285847664, + 0.1356714963912964, + 0.07641928642988205, + 0.12885114550590515, + 0.03607892990112305, + 0.04021726921200752, + 0.1383206844329834, + 0.11742430180311203 + ], + "credit_map": [ + 0.022116757929325104, + 0.020749537274241447, + 0.009517847560346127, + 0.010341052897274494, + 0.03835979849100113, + 0.01627650484442711, + 0.018693042919039726, + 0.013246063143014908, + 0.033386506140232086, + 0.03163592517375946, + 0.0068421331234276295, + 0.007617671974003315, + 0.013511301018297672, + 0.020261140540242195, + 0.004576853010803461, + 0.006277301348745823, + 0.009989108890295029, + 0.00729368906468153, + 0.01239597238600254, + 0.005211560055613518, + 0.005035178270190954, + 0.006284930743277073, + 0.0033904199954122305, + 0.0040065632201731205, + 0.02350340411067009, + 0.019747108221054077, + 0.014717071317136288, + 0.016473228111863136, + 0.011112279258668423, + 0.021093161776661873, + 0.012769509106874466, + 0.016255738213658333, + 0.00660958606749773, + 0.005841346457600594, + 0.023312000557780266, + 0.004580180160701275, + 0.00942982267588377, + 0.0058257607743144035, + 0.0038500025402754545, + 0.016970593482255936, + 0.03228331357240677, + 0.012113088741898537, + 0.011553142219781876, + 0.008594349026679993, + 0.006182012148201466, + 0.00853619072586298, + 0.010399345308542252, + 0.039189692586660385, + 0.003767674556002021, + 0.005878068041056395, + 0.0036409632302820683, + 0.0038494165055453777, + 0.00519115524366498, + 0.0050852736458182335, + 0.004808421712368727, + 0.0038579574320465326, + 0.006116450764238834, + 0.004950086120516062, + 0.008096633478999138, + 0.003996363840997219, + 0.0047609019093215466, + 0.004424979444593191, + 0.003112275153398514, + 0.004759577568620443, + 0.04881074279546738, + 0.053618621081113815, + 0.005412154830992222, + 0.003324313322082162, + 0.011435950174927711, + 0.008556145243346691, + 0.003448118455708027, + 0.00371464341878891, + 0.011798826977610588, + 0.02389831840991974, + 0.01695391535758972, + 0.01773517206311226, + 0.012534482404589653, + 0.01331852562725544, + 0.010268021374940872, + 0.010917051695287228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.07004504650831223 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.004638507030904293 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638507030904293, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638507030904293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759454309940338, + 0.0726856216788292, + 0.04864862188696861, + 0.049594614654779434, + 0.15915410220623016, + 0.0778781920671463, + 0.06305188685655594, + 0.0800393596291542, + 0.15329109132289886, + 0.11971107870340347 + ], + "credit_map": [ + 0.02478465624153614, + 0.02678079530596733, + 0.0151760159060359, + 0.017862459644675255, + 0.03658691421151161, + 0.02091986872255802, + 0.015509716235101223, + 0.018325012177228928, + 0.011639389209449291, + 0.0122558968141675, + 0.00484453747048974, + 0.00727793900296092, + 0.01550314575433731, + 0.012680646032094955, + 0.004867952316999435, + 0.0036161195021122694, + 0.007082117255777121, + 0.007430655881762505, + 0.013141686096787453, + 0.005504656583070755, + 0.003987674601376057, + 0.0046224468387663364, + 0.0025086000096052885, + 0.004370782990008593, + 0.012162444181740284, + 0.014973490498960018, + 0.00301022594794631, + 0.0030905622988939285, + 0.00536127807572484, + 0.003973667044192553, + 0.003893462708219886, + 0.0031294801738113165, + 0.014039967209100723, + 0.010262801311910152, + 0.05693568289279938, + 0.008268304169178009, + 0.013214276172220707, + 0.008301413618028164, + 0.007625107187777758, + 0.04050655663013458, + 0.016375308856368065, + 0.006417025811970234, + 0.007707942742854357, + 0.0057396418415009975, + 0.005041578318923712, + 0.004890351090580225, + 0.008753468282520771, + 0.022952867671847343, + 0.007792243734002113, + 0.007010415196418762, + 0.005523549858480692, + 0.004509073216468096, + 0.008864647708833218, + 0.01770753040909767, + 0.004520289599895477, + 0.007124137133359909, + 0.011030321009457111, + 0.012061686255037785, + 0.009988555684685707, + 0.009023954160511494, + 0.010826360434293747, + 0.009528204798698425, + 0.011951856315135956, + 0.005628417246043682, + 0.055461809039115906, + 0.054208073765039444, + 0.007952621206641197, + 0.005080766044557095, + 0.012167911045253277, + 0.009909852407872677, + 0.003910598810762167, + 0.004599468316882849, + 0.010393543168902397, + 0.02213330566883087, + 0.019879158586263657, + 0.01983601786196232, + 0.013314890675246716, + 0.013890016824007034, + 0.00906217098236084, + 0.011201979592442513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.004638507030904293 + }, + { + "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, minimalist stainless steel ballpoint pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen moving in a perfectly elliptical trajectory through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a studio with warm golden light filtering through a large window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred scene of carpentry tools, wood chips, and shelves creating a soft bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly while keeping the pen perfectly centered and in sharp 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": [ + -0.2521213889122009 + ], + "opposite_logprob_tokens": [ + -1.5021214485168457 + ], + "logprob_reward": -0.2521213889122009, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521213889122009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07230957597494125, + 0.0266561396420002, + 0.025053206831216812, + 0.026769466698169708, + 0.021140214055776596, + 0.3513028025627136, + 0.07097161561250687, + 0.1253083050251007, + 0.2216237187385559, + 0.05886494368314743 + ], + "credit_map": [ + 0.009907777421176434, + 0.013691571541130543, + 0.005179483909159899, + 0.0058873617090284824, + 0.015436522662639618, + 0.008480178192257881, + 0.006457568146288395, + 0.007269116584211588, + 0.008050468750298023, + 0.007304089143872261, + 0.0012865783646702766, + 0.0014507310697808862, + 0.0025855419225990772, + 0.0033822793047875166, + 0.001286685117520392, + 0.0013097660848870873, + 0.0042174868285655975, + 0.0034960161428898573, + 0.003412742167711258, + 0.003324240678921342, + 0.002731266664341092, + 0.0030821161344647408, + 0.0015001362189650536, + 0.003289200132712722, + 0.007641368079930544, + 0.008309155702590942, + 0.0011724228970706463, + 0.0016028434038162231, + 0.0027191804256290197, + 0.002151908352971077, + 0.002026212401688099, + 0.0011463768314570189, + 0.002605941379442811, + 0.0021181947086006403, + 0.0034959488548338413, + 0.001347677898593247, + 0.004232748411595821, + 0.002071494236588478, + 0.0029501470271497965, + 0.0023180623538792133, + 0.09118558466434479, + 0.029815947636961937, + 0.03137246519327164, + 0.017810886725783348, + 0.013709700666368008, + 0.022063056007027626, + 0.02533770352602005, + 0.12000741809606552, + 0.0063805957324802876, + 0.00811052042990923, + 0.004999638069421053, + 0.007425088435411453, + 0.008595120161771774, + 0.017673509195446968, + 0.008286171592772007, + 0.009500972926616669, + 0.017057478427886963, + 0.021680200472474098, + 0.015609742142260075, + 0.011745668947696686, + 0.018859798088669777, + 0.013388950377702713, + 0.01726711541414261, + 0.009699352085590363, + 0.0769931823015213, + 0.08562490344047546, + 0.008944368921220303, + 0.004756034351885319, + 0.018931357190012932, + 0.014541939832270145, + 0.005692832171916962, + 0.006139103788882494, + 0.0056590549647808075, + 0.009724930860102177, + 0.0056162625551223755, + 0.007746560033410788, + 0.007372750900685787, + 0.01009088009595871, + 0.0058987559750676155, + 0.0067557464353740215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.2521213889122009 + } + ], + "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/0000167/group_15/index_tensor([8592, 8592])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_tensor([8592, 8592])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..130940b9b234a6aa941db51139a6ddb290b2cb07 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_tensor([8592, 8592])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tilted Dutch-angle wide-angle lens, with an unsteady tracking motion, captures a surreal stroll through a city late at night. A massive, bioluminescent siphonophore, the size of a parade float, floats slowly and majestically in the center of a deserted street. Its translucent, gelatinous body, composed of countless tiny, interconnected individuals, emanates a cold blue and purple glow, a stark contrast to the warm, flickering neon lights of its surroundings. Its massive, umbrella-shaped propulsion organs rhythmically contract and expand, each beat pulsing outward like a heart, a brilliant cyan halo and accompanied by a clear, ethereal cry that echoes between the towering skyscrapers, whose windows are mostly darkened. The damp asphalt reflects the colorful lights of the signs and the luminescent glow of the strange creature itself like a mirror. Countless luminescent tentacles dangle from its lower body like ornate tassels, swaying gently, occasionally skimming the ground, leaving fleeting traces of phosphorescence on the damp pavement. The entire picture appears both dreamy and unsettling due to the tilted composition, creating a grand, lonely and awe-inspiring visual spectacle.", + "seed": 31850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_8592_seed_31850_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": -0.016777609754353762, + "vlm_logprob": -0.016777609754353762, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.016777609754353762, + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01427921187132597 + ], + "opposite_logprob_tokens": [ + -4.514279365539551 + ], + "logprob_reward": -0.01427921187132597, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01427921187132597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07427602261304855, + 0.4918067455291748, + 0.08920798450708389, + 0.032086081802845, + 0.055177122354507446, + 0.033941298723220825, + 0.03336960822343826, + 0.08123970031738281, + 0.040308330208063126, + 0.06858711689710617 + ], + "credit_map": [ + 0.011477882042527199, + 0.007117639761418104, + 0.015457616187632084, + 0.006336083635687828, + 0.008581746369600296, + 0.0060549029149115086, + 0.005621371790766716, + 0.013628772459924221, + 0.046382609754800797, + 0.07609278708696365, + 0.046640798449516296, + 0.05163310840725899, + 0.1475590616464615, + 0.030185705050826073, + 0.05801719054579735, + 0.03529549762606621, + 0.01352628506720066, + 0.007567990571260452, + 0.01581604778766632, + 0.0069632139056921005, + 0.012247495353221893, + 0.011138512752950191, + 0.008206745609641075, + 0.013741688802838326, + 0.005519072059541941, + 0.002582421526312828, + 0.00521157868206501, + 0.0021169008687138557, + 0.0037339916452765465, + 0.006138207390904427, + 0.002982953330501914, + 0.003800956066697836, + 0.006459723226726055, + 0.003269356908276677, + 0.013116893358528614, + 0.00522998021915555, + 0.006629705894738436, + 0.003663969924673438, + 0.0044702948071062565, + 0.012337199412286282, + 0.004166137892752886, + 0.0028670569881796837, + 0.008189883083105087, + 0.0032834215089678764, + 0.0028841884341090918, + 0.00322724599391222, + 0.0021147113293409348, + 0.0072086527943611145, + 0.0032943938858807087, + 0.003767473855987191, + 0.0045398869551718235, + 0.003996171522885561, + 0.006359357852488756, + 0.003921151626855135, + 0.0031996057368814945, + 0.004291565623134375, + 0.007158280350267887, + 0.013702579773962498, + 0.007708868943154812, + 0.00980893149971962, + 0.014465943910181522, + 0.00803239457309246, + 0.010419405996799469, + 0.00994329247623682, + 0.005358222406357527, + 0.0037854579277336597, + 0.0052708713337779045, + 0.006339824292808771, + 0.004635644610971212, + 0.003072087187319994, + 0.006199260707944632, + 0.005646963138133287, + 0.011004909873008728, + 0.008303951472043991, + 0.01309641171246767, + 0.008750034496188164, + 0.007625692058354616, + 0.005102419760078192, + 0.006254524923861027, + 0.008449173532426357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.01427921187132597 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011887634173035622 + ], + "opposite_logprob_tokens": [ + -6.7511887550354 + ], + "logprob_reward": -0.0011887634173035622, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011887634173035622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2669152319431305, + 0.09419361501932144, + 0.052381835877895355, + 0.05577641725540161, + 0.15861544013023376, + 0.057679735124111176, + 0.02961074560880661, + 0.16815327107906342, + 0.053171947598457336, + 0.0635017603635788 + ], + "credit_map": [ + 0.037308625876903534, + 0.026797758415341377, + 0.07906388491392136, + 0.014330851845443249, + 0.021584151312708855, + 0.013258443213999271, + 0.014484437182545662, + 0.060087066143751144, + 0.008976675570011139, + 0.01432211697101593, + 0.008233270607888699, + 0.009839445352554321, + 0.02619359828531742, + 0.00634436309337616, + 0.012813654728233814, + 0.007470495067536831, + 0.00988274160772562, + 0.004510064143687487, + 0.007365921977907419, + 0.004945025313645601, + 0.006023549009114504, + 0.008046538569033146, + 0.004283287562429905, + 0.007324711885303259, + 0.008305065333843231, + 0.003851631423458457, + 0.010899065062403679, + 0.0027117172721773386, + 0.0056241946294903755, + 0.007340727839618921, + 0.004385641310364008, + 0.012658374384045601, + 0.014847136102616787, + 0.012183052487671375, + 0.033061474561691284, + 0.011471922509372234, + 0.02711886540055275, + 0.01723284088075161, + 0.010609236545860767, + 0.032090924680233, + 0.006538258399814367, + 0.006355460733175278, + 0.013693094253540039, + 0.00544392503798008, + 0.005878508556634188, + 0.0042267232201993465, + 0.004195626825094223, + 0.011348138563334942, + 0.0039050474297255278, + 0.0029457188211381435, + 0.004240317735821009, + 0.0035913323517888784, + 0.004554033279418945, + 0.0026799235492944717, + 0.003556199138984084, + 0.0041381726041436195, + 0.010541611351072788, + 0.0387825109064579, + 0.009864602237939835, + 0.01489392388612032, + 0.03977396339178085, + 0.014033830724656582, + 0.02174205146729946, + 0.018520774319767952, + 0.003497194731608033, + 0.008775065653026104, + 0.004501744173467159, + 0.0058072712272405624, + 0.011282532475888729, + 0.005582763347774744, + 0.007442439440637827, + 0.006282940041273832, + 0.005651043262332678, + 0.008981899358332157, + 0.006623379420489073, + 0.007675804663449526, + 0.013260718435049057, + 0.005419399123638868, + 0.008814215660095215, + 0.00707530602812767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0011887634173035622 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.20791931450366974, + 0.1774531751871109, + 0.05037744343280792, + 0.04215453192591667, + 0.23518076539039612, + 0.03518659994006157, + 0.028914673253893852, + 0.04034804925322533, + 0.0631374716758728, + 0.11932791024446487 + ], + "credit_map": [ + 0.026699796319007874, + 0.021158767864108086, + 0.056412771344184875, + 0.010969223454594612, + 0.020444272086024284, + 0.011385343968868256, + 0.012261982075870037, + 0.04858715832233429, + 0.017096929252147675, + 0.027301454916596413, + 0.016709808260202408, + 0.01812988705933094, + 0.05386791005730629, + 0.010747510939836502, + 0.020934492349624634, + 0.012665185146033764, + 0.008013549260795116, + 0.004887489136308432, + 0.007012302055954933, + 0.005919755436480045, + 0.00781975220888853, + 0.0043974569998681545, + 0.004450780339539051, + 0.00787635613232851, + 0.006489689461886883, + 0.0025705737061798573, + 0.0080903684720397, + 0.002888820134103298, + 0.0035357701126486063, + 0.005291470792144537, + 0.0034072885755449533, + 0.009880551137030125, + 0.022075902670621872, + 0.01123854424804449, + 0.07812894135713577, + 0.012763168662786484, + 0.018216032534837723, + 0.010654461570084095, + 0.010006540454924107, + 0.07209718972444534, + 0.003648904850706458, + 0.0034216621425002813, + 0.0082129817456007, + 0.00320375245064497, + 0.0041035315953195095, + 0.0024940799921751022, + 0.0031847329810261726, + 0.00691695511341095, + 0.004093743860721588, + 0.0025740363635122776, + 0.004365265369415283, + 0.0033332218881696463, + 0.004125404637306929, + 0.002905247500166297, + 0.0031529979314655066, + 0.004364758729934692, + 0.004122002515941858, + 0.0059717874974012375, + 0.003792840288951993, + 0.004341918043792248, + 0.005934986285865307, + 0.004798995330929756, + 0.006763448007404804, + 0.004622071050107479, + 0.0066057960502803326, + 0.012777319177985191, + 0.005956438370049, + 0.005291532725095749, + 0.012088857591152191, + 0.006152618210762739, + 0.006901591084897518, + 0.00736332219094038, + 0.007784421090036631, + 0.015796802937984467, + 0.014762931503355503, + 0.021961303427815437, + 0.023138046264648438, + 0.008668210357427597, + 0.012598972767591476, + 0.014617228880524635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.006735715549439192 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494414269924164 + ], + "opposite_logprob_tokens": [ + -2.9299440383911133 + ], + "logprob_reward": -0.05494414269924164, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494414269924164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14988423883914948, + 0.20468057692050934, + 0.03484996035695076, + 0.07196848094463348, + 0.20862315595149994, + 0.05255847051739693, + 0.060646191239356995, + 0.07319614291191101, + 0.023422090336680412, + 0.12017069011926651 + ], + "credit_map": [ + 0.019629735499620438, + 0.011887793429195881, + 0.04046536609530449, + 0.008620868436992168, + 0.011279728263616562, + 0.011853712610900402, + 0.010209159925580025, + 0.03593787923455238, + 0.01828109100461006, + 0.02989613078534603, + 0.018739687278866768, + 0.021719684824347496, + 0.06004483625292778, + 0.014392091892659664, + 0.02574921026825905, + 0.01585785672068596, + 0.004177914932370186, + 0.00423562154173851, + 0.005147707648575306, + 0.004744356498122215, + 0.005681538954377174, + 0.004031577613204718, + 0.0033359359949827194, + 0.0034953069407492876, + 0.012159151956439018, + 0.004849561024457216, + 0.011984897777438164, + 0.005198588594794273, + 0.004817648325115442, + 0.011158222332596779, + 0.0067209345288574696, + 0.015079478733241558, + 0.02022106759250164, + 0.008056453429162502, + 0.06700127571821213, + 0.012610197998583317, + 0.015321025624871254, + 0.011979186907410622, + 0.009117109701037407, + 0.06431683897972107, + 0.00597395421937108, + 0.009181179106235504, + 0.007888555526733398, + 0.005253326613456011, + 0.007890061475336552, + 0.004143996629863977, + 0.0046522971242666245, + 0.007575099356472492, + 0.007555450312793255, + 0.006495897192507982, + 0.007739249616861343, + 0.01017853058874607, + 0.006676806602627039, + 0.0063972510397434235, + 0.0069856466725468636, + 0.008617358282208443, + 0.005240770056843758, + 0.01566394791007042, + 0.005955126136541367, + 0.00872783362865448, + 0.017275037243962288, + 0.006414636969566345, + 0.006604163441807032, + 0.007314622867852449, + 0.00205681542865932, + 0.003064206801354885, + 0.0023623693268746138, + 0.003195028053596616, + 0.0035692360252141953, + 0.002214280888438225, + 0.0037530602421611547, + 0.00320709147490561, + 0.008051306940615177, + 0.023436853662133217, + 0.01012632716447115, + 0.008549753576517105, + 0.0327560119330883, + 0.00847291573882103, + 0.014926197938621044, + 0.013851312920451164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.05494414269924164 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067402152344584465 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.0067402152344584465, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067402152344584465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23429203033447266, + 0.4899402856826782, + 0.04022680222988129, + 0.022546416148543358, + 0.08808380365371704, + 0.02455681376159191, + 0.009931315667927265, + 0.019307522103190422, + 0.040627650916576385, + 0.030487380921840668 + ], + "credit_map": [ + 0.03722270950675011, + 0.015691939741373062, + 0.07645299285650253, + 0.0119321895763278, + 0.013846848160028458, + 0.011782106943428516, + 0.011725934222340584, + 0.055637311190366745, + 0.045938730239868164, + 0.07528254389762878, + 0.04627474769949913, + 0.051404401659965515, + 0.14815694093704224, + 0.03154192492365837, + 0.05735621228814125, + 0.03398476913571358, + 0.0057271015830338, + 0.004013936035335064, + 0.005255316849797964, + 0.002773368265479803, + 0.006677653174847364, + 0.0057761818170547485, + 0.0056180888786911964, + 0.004385150503367186, + 0.0039035372901707888, + 0.001790160546079278, + 0.005158794112503529, + 0.0016121083172038198, + 0.0022745782043784857, + 0.0031324592418968678, + 0.0016392201650887728, + 0.0030355555936694145, + 0.009363891556859016, + 0.006849116645753384, + 0.01950130984187126, + 0.009938816539943218, + 0.010910376906394958, + 0.00666520232334733, + 0.008118017576634884, + 0.016737058758735657, + 0.0018695777980610728, + 0.004000215325504541, + 0.004131953231990337, + 0.0026601627469062805, + 0.003932904452085495, + 0.0019706289749592543, + 0.0026826313696801662, + 0.0033087371848523617, + 0.001266488921828568, + 0.000806151598226279, + 0.001616748864762485, + 0.0013435864821076393, + 0.0010071499273180962, + 0.0013337030541151762, + 0.0010534652974456549, + 0.0015040204161778092, + 0.0020775909069925547, + 0.0032201046124100685, + 0.0021085988264530897, + 0.0022571301087737083, + 0.0027993645053356886, + 0.0022091541904956102, + 0.0023875595070421696, + 0.0022480173502117395, + 0.003109729615971446, + 0.005127847660332918, + 0.004105099942535162, + 0.007543738000094891, + 0.005240990314632654, + 0.0036528627388179302, + 0.00597888370975852, + 0.005868491716682911, + 0.0033474520314484835, + 0.003633739659562707, + 0.004763495642691851, + 0.004765369929373264, + 0.004321769345551729, + 0.002354556927457452, + 0.003399993758648634, + 0.0039009987376630306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0067402152344584465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted Dutch-angle wide-angle lens, with an unsteady tracking motion, captures a surreal stroll through a city late at night. A massive, bioluminescent siphonophore, the size of a parade float, floats slowly and majestically in the center of a deserted street. Its translucent, gelatinous body, composed of countless tiny, interconnected individuals, emanates a cold blue and purple glow, a stark contrast to the warm, flickering neon lights of its surroundings. Its massive, umbrella-shaped propulsion organs rhythmically contract and expand, each beat pulsing outward like a heart, a brilliant cyan halo and accompanied by a clear, ethereal cry that echoes between the towering skyscrapers, whose windows are mostly darkened. The damp asphalt reflects the colorful lights of the signs and the luminescent glow of the strange creature itself like a mirror. Countless luminescent tentacles dangle from its lower body like ornate tassels, swaying gently, occasionally skimming the ground, leaving fleeting traces of phosphorescence on the damp pavement. The entire picture appears both dreamy and unsettling due to the tilted composition, creating a grand, lonely and awe-inspiring visual spectacle.", + "seed": 31851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_8592_seed_31851_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": -0.017584300227463247, + "vlm_logprob": -0.017584300227463247, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.017584300227463247, + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04378798231482506 + ], + "opposite_logprob_tokens": [ + -3.168787956237793 + ], + "logprob_reward": -0.04378798231482506, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04378798231482506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044435054063797, + 0.10402335971593857, + 0.22200776636600494, + 0.09036368131637573, + 0.05141949653625488, + 0.15079477429389954, + 0.03348373621702194, + 0.03611323609948158, + 0.1399284303188324, + 0.067422054708004 + ], + "credit_map": [ + 0.012824173085391521, + 0.012377247214317322, + 0.013414655812084675, + 0.015692513436079025, + 0.008662265725433826, + 0.013555736280977726, + 0.013771395199000835, + 0.014145512133836746, + 0.0069455415941774845, + 0.007158389315009117, + 0.010323414579033852, + 0.012743464671075344, + 0.010143903084099293, + 0.033051516860723495, + 0.011863935738801956, + 0.011793188750743866, + 0.02799336239695549, + 0.01907331496477127, + 0.03981135040521622, + 0.026268815621733665, + 0.023242846131324768, + 0.03949028253555298, + 0.022807085886597633, + 0.023320699110627174, + 0.0062596239149570465, + 0.01078438013792038, + 0.0068835848942399025, + 0.012556982226669788, + 0.012833992019295692, + 0.017321694642305374, + 0.011068102903664112, + 0.012655314058065414, + 0.0034451468382030725, + 0.004184254445135593, + 0.005072594620287418, + 0.0059744855388998985, + 0.006027545314282179, + 0.011794875375926495, + 0.0057800947688519955, + 0.009140495210886002, + 0.016400950029492378, + 0.01425638422369957, + 0.018065396696329117, + 0.01801828108727932, + 0.020986486226320267, + 0.031186653301119804, + 0.014577638357877731, + 0.01730298064649105, + 0.0036704542580991983, + 0.005275679286569357, + 0.004709511063992977, + 0.004767652601003647, + 0.0040437825955450535, + 0.0034817170817404985, + 0.0035702900495380163, + 0.003964647650718689, + 0.00547153502702713, + 0.005575172603130341, + 0.0047625210136175156, + 0.004760542418807745, + 0.004397185519337654, + 0.0037634545005857944, + 0.0031789273489266634, + 0.0042038969695568085, + 0.018401801586151123, + 0.0201101116836071, + 0.022386783733963966, + 0.015450947917997837, + 0.011996731162071228, + 0.020156100392341614, + 0.017525214701890945, + 0.013900717720389366, + 0.005394237115979195, + 0.006316867657005787, + 0.008437934331595898, + 0.007362554781138897, + 0.005146559327840805, + 0.016119102016091347, + 0.01009298488497734, + 0.008551809936761856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.04378798231482506 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007250064518302679 + ], + "opposite_logprob_tokens": [ + -7.250724792480469 + ], + "logprob_reward": -0.0007250064518302679, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007250064518302679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09063859283924103, + 0.11073729395866394, + 0.1468949168920517, + 0.043504003435373306, + 0.08925693482160568, + 0.06336517632007599, + 0.13411222398281097, + 0.07928904891014099, + 0.18586087226867676, + 0.05634089186787605 + ], + "credit_map": [ + 0.010621913708746433, + 0.014564834535121918, + 0.006839873269200325, + 0.011194977909326553, + 0.011114685796201229, + 0.009322606958448887, + 0.01800604537129402, + 0.008973650634288788, + 0.010865436866879463, + 0.008380780927836895, + 0.014117066748440266, + 0.013003422878682613, + 0.0118956808000803, + 0.03177386894822121, + 0.009571735747158527, + 0.011129303835332394, + 0.017669934779405594, + 0.011640986427664757, + 0.022033890709280968, + 0.021476680412888527, + 0.016831493005156517, + 0.027152743190526962, + 0.016711704432964325, + 0.013377482071518898, + 0.002343868836760521, + 0.0034583560191094875, + 0.003991159610450268, + 0.005509528331458569, + 0.006919152103364468, + 0.008659060113132, + 0.006556822918355465, + 0.006066052708774805, + 0.0067871371284127235, + 0.006441105157136917, + 0.010767187923192978, + 0.00972865242511034, + 0.010571508668363094, + 0.020669320598244667, + 0.008336271159350872, + 0.015955748036503792, + 0.006160684861242771, + 0.007568903733044863, + 0.00936554279178381, + 0.008189723826944828, + 0.008693402633070946, + 0.010734089650213718, + 0.005844708066433668, + 0.006808117963373661, + 0.024997856467962265, + 0.0190906822681427, + 0.024607907980680466, + 0.011710652150213718, + 0.01624455861747265, + 0.01190322544425726, + 0.010175942443311214, + 0.015381395816802979, + 0.010246949270367622, + 0.007553054951131344, + 0.01218858640640974, + 0.013871685601770878, + 0.009332126937806606, + 0.005629810970276594, + 0.00686552282422781, + 0.013601317070424557, + 0.023925021290779114, + 0.016386743634939194, + 0.027295291423797607, + 0.02550312876701355, + 0.019123904407024384, + 0.028407003730535507, + 0.0243275985121727, + 0.020892193540930748, + 0.0055511631071567535, + 0.005921189207583666, + 0.007298727985471487, + 0.005725027061998844, + 0.004538234323263168, + 0.010146960616111755, + 0.006348071154206991, + 0.01081151980906725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.0007250064518302679 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028191369492560625 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028191369492560625, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028191369492560625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08338922262191772, + 0.04833747446537018, + 0.06928140670061111, + 0.03361973166465759, + 0.1997772455215454, + 0.0744701623916626, + 0.07449961453676224, + 0.18468649685382843, + 0.18080134689807892, + 0.05113725736737251 + ], + "credit_map": [ + 0.008786552585661411, + 0.011113052256405354, + 0.009487367235124111, + 0.013203708454966545, + 0.010450274683535099, + 0.008280380629003048, + 0.012873890809714794, + 0.009193998761475086, + 0.004524485673755407, + 0.0033372954931110144, + 0.0063019245862960815, + 0.0062799155712127686, + 0.005426338873803616, + 0.011259487830102444, + 0.005025727208703756, + 0.006182300392538309, + 0.007083302829414606, + 0.006186134181916714, + 0.01167878694832325, + 0.012294298969209194, + 0.005563716404139996, + 0.011965633369982243, + 0.008590114302933216, + 0.0059194182977080345, + 0.003285233397036791, + 0.004202241078019142, + 0.0037789943162351847, + 0.004226278513669968, + 0.004150973632931709, + 0.006099841091781855, + 0.0034391903318464756, + 0.004436980001628399, + 0.014470595866441727, + 0.014162419363856316, + 0.024596644565463066, + 0.022304445505142212, + 0.021354930475354195, + 0.04376072436571121, + 0.019666923210024834, + 0.039460569620132446, + 0.00780915655195713, + 0.00710642896592617, + 0.009336194023489952, + 0.009257053956389427, + 0.011231805197894573, + 0.014885025098919868, + 0.006674232427030802, + 0.008170265704393387, + 0.01212663296610117, + 0.009969567880034447, + 0.012879944406449795, + 0.007429187186062336, + 0.00853927806019783, + 0.007294321432709694, + 0.0069106947630643845, + 0.009349989704787731, + 0.028700290247797966, + 0.01821731962263584, + 0.031949326395988464, + 0.028632625937461853, + 0.019018037244677544, + 0.015203362330794334, + 0.01558049488812685, + 0.0273850429803133, + 0.016452910378575325, + 0.038422416895627975, + 0.019385116174817085, + 0.022525208070874214, + 0.030159026384353638, + 0.017252005636692047, + 0.021948309615254402, + 0.014656351879239082, + 0.004662638530135155, + 0.006893354933708906, + 0.006216703448444605, + 0.005021306686103344, + 0.005676849279552698, + 0.008103428408503532, + 0.007700292393565178, + 0.00686268275603652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0028191369492560625 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13696150481700897, + 0.055148277431726456, + 0.1844501495361328, + 0.12848308682441711, + 0.1072765663266182, + 0.14080631732940674, + 0.04266401380300522, + 0.053396906703710556, + 0.10102161765098572, + 0.049791619181632996 + ], + "credit_map": [ + 0.013011952862143517, + 0.011651607230305672, + 0.020189018920063972, + 0.021180829033255577, + 0.013552244752645493, + 0.015868527814745903, + 0.022211873903870583, + 0.019295429810881615, + 0.003915837034583092, + 0.004488095641136169, + 0.005606533959507942, + 0.006439079064875841, + 0.0064599947072565556, + 0.013561192899942398, + 0.005966890137642622, + 0.008710653521120548, + 0.021617570891976357, + 0.015903549268841743, + 0.03526659682393074, + 0.024460088461637497, + 0.01631847769021988, + 0.03376014530658722, + 0.019503943622112274, + 0.017619768157601357, + 0.008677749894559383, + 0.017877858132123947, + 0.010290359146893024, + 0.0210580937564373, + 0.016752561554312706, + 0.021630337461829185, + 0.01671753078699112, + 0.015478594228625298, + 0.005848911125212908, + 0.008045407943427563, + 0.009010117501020432, + 0.010506800375878811, + 0.012849977239966393, + 0.030489858239889145, + 0.011868002824485302, + 0.018657494336366653, + 0.018089959397912025, + 0.017391502857208252, + 0.017385665327310562, + 0.014748056419193745, + 0.01891564391553402, + 0.025760389864444733, + 0.01273181289434433, + 0.0157832782715559, + 0.005229865200817585, + 0.007124711759388447, + 0.005501207895576954, + 0.004718095529824495, + 0.00502618495374918, + 0.005090932361781597, + 0.004158497788012028, + 0.005814515054225922, + 0.0077987597323954105, + 0.005594127811491489, + 0.00856178067624569, + 0.007298690266907215, + 0.005469809751957655, + 0.005424367263913155, + 0.0051693981513381, + 0.008079967461526394, + 0.012014265172183514, + 0.01663948781788349, + 0.013301958329975605, + 0.01137456577271223, + 0.013164181262254715, + 0.011827285401523113, + 0.013223381713032722, + 0.009476488456130028, + 0.0056167286820709705, + 0.003912958782166243, + 0.0055144778452813625, + 0.006447703577578068, + 0.003946180921047926, + 0.009255185723304749, + 0.006564082112163305, + 0.008534295484423637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03809695690870285 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09574156999588013, + 0.10609167814254761, + 0.11733565479516983, + 0.058165933936834335, + 0.02514580637216568, + 0.22496482729911804, + 0.043968625366687775, + 0.05332174897193909, + 0.20482981204986572, + 0.07043442875146866 + ], + "credit_map": [ + 0.008658578619360924, + 0.013233177363872528, + 0.011168439872562885, + 0.016465933993458748, + 0.010203377343714237, + 0.011229606345295906, + 0.01375536248087883, + 0.011027072556316853, + 0.010009551420807838, + 0.007723273243755102, + 0.014268367551267147, + 0.011750703677535057, + 0.012990984134376049, + 0.025780798867344856, + 0.011479503475129604, + 0.012088475748896599, + 0.015266725793480873, + 0.009998857043683529, + 0.02160431258380413, + 0.015729019418358803, + 0.01033521257340908, + 0.021452080458402634, + 0.012014954350888729, + 0.010934473015367985, + 0.0038574545178562403, + 0.005880925338715315, + 0.006876610219478607, + 0.008142691105604172, + 0.009251058101654053, + 0.009930736385285854, + 0.00675398763269186, + 0.007472455035895109, + 0.0026014174800366163, + 0.0021628476679325104, + 0.0027678031474351883, + 0.002891227602958679, + 0.0024128712248057127, + 0.005749893374741077, + 0.0024801308754831553, + 0.0040796115063130856, + 0.027815837413072586, + 0.020108817145228386, + 0.029889622703194618, + 0.024085121229290962, + 0.02756522409617901, + 0.04842314496636391, + 0.02034948766231537, + 0.026727542281150818, + 0.007583031430840492, + 0.006254847161471844, + 0.007775825913995504, + 0.004632278811186552, + 0.0045967865735292435, + 0.004120273515582085, + 0.004017559811472893, + 0.0049880133010447025, + 0.007403504103422165, + 0.005796639248728752, + 0.008434480056166649, + 0.009401656687259674, + 0.004942028317600489, + 0.005440517794340849, + 0.005196630489081144, + 0.006706282030791044, + 0.02152668684720993, + 0.05099630728363991, + 0.027677973732352257, + 0.020648464560508728, + 0.03495235741138458, + 0.01699836552143097, + 0.017961567267775536, + 0.01406804472208023, + 0.0066956221126019955, + 0.00597978662699461, + 0.010198630392551422, + 0.007070244289934635, + 0.005660414230078459, + 0.016784919425845146, + 0.00879102572798729, + 0.009253778494894505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.0024924185127019882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted Dutch-angle wide-angle lens, with an unsteady tracking motion, captures a surreal stroll through a city late at night. A massive, bioluminescent siphonophore, the size of a parade float, floats slowly and majestically in the center of a deserted street. Its translucent, gelatinous body, composed of countless tiny, interconnected individuals, emanates a cold blue and purple glow, a stark contrast to the warm, flickering neon lights of its surroundings. Its massive, umbrella-shaped propulsion organs rhythmically contract and expand, each beat pulsing outward like a heart, a brilliant cyan halo and accompanied by a clear, ethereal cry that echoes between the towering skyscrapers, whose windows are mostly darkened. The damp asphalt reflects the colorful lights of the signs and the luminescent glow of the strange creature itself like a mirror. Countless luminescent tentacles dangle from its lower body like ornate tassels, swaying gently, occasionally skimming the ground, leaving fleeting traces of phosphorescence on the damp pavement. The entire picture appears both dreamy and unsettling due to the tilted composition, creating a grand, lonely and awe-inspiring visual spectacle.", + "seed": 31852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_8592_seed_31852_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": -0.0571172708994709, + "vlm_logprob": -0.0571172708994709, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.0571172708994709, + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11506136506795883 + ], + "opposite_logprob_tokens": [ + -2.2400612831115723 + ], + "logprob_reward": -0.11506136506795883, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11506136506795883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15884342789649963, + 0.15435612201690674, + 0.11993519216775894, + 0.05974215269088745, + 0.07999183982610703, + 0.05365104600787163, + 0.16253723204135895, + 0.04349302873015404, + 0.09927357733249664, + 0.06817643344402313 + ], + "credit_map": [ + 0.01520423125475645, + 0.020889682695269585, + 0.02198321372270584, + 0.020013611763715744, + 0.023345505818724632, + 0.011999213136732578, + 0.026079086586833, + 0.01932889223098755, + 0.01291666179895401, + 0.020119015127420425, + 0.022679051384329796, + 0.014871385879814625, + 0.0322551429271698, + 0.016991212964057922, + 0.015892740339040756, + 0.018630890175700188, + 0.009771709330379963, + 0.012548306956887245, + 0.014913497492671013, + 0.0112508749589324, + 0.03623100370168686, + 0.010644825175404549, + 0.013390105217695236, + 0.011184859089553356, + 0.008958634920418262, + 0.004685680847615004, + 0.010820181109011173, + 0.004594690166413784, + 0.006695470307022333, + 0.008510923013091087, + 0.003994429018348455, + 0.011482144705951214, + 0.005973683204501867, + 0.011474336497485638, + 0.00855143554508686, + 0.01312586385756731, + 0.014988992363214493, + 0.005739258136600256, + 0.011864897795021534, + 0.00827337708324194, + 0.005095789674669504, + 0.005551690235733986, + 0.008105079643428326, + 0.008435243740677834, + 0.007023607846349478, + 0.006732418667525053, + 0.006370089948177338, + 0.006337120663374662, + 0.010646392591297626, + 0.017254561185836792, + 0.01576685719192028, + 0.0329907201230526, + 0.0255909264087677, + 0.011943802237510681, + 0.02673269994556904, + 0.021611269563436508, + 0.0037214390467852354, + 0.003642754163593054, + 0.0041505638509988785, + 0.009281439706683159, + 0.00566494045779109, + 0.005067160353064537, + 0.005928827915340662, + 0.006035900209099054, + 0.010227949358522892, + 0.01623787172138691, + 0.01549977995455265, + 0.010989789851009846, + 0.014124434441328049, + 0.009175501763820648, + 0.008002535440027714, + 0.01501570176333189, + 0.015273131430149078, + 0.004581107292324305, + 0.008389503695070744, + 0.007862427271902561, + 0.008913399651646614, + 0.008093373849987984, + 0.007053243927657604, + 0.008010241203010082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.11506136506795883 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008219677838496864 + ], + "opposite_logprob_tokens": [ + -7.125822067260742 + ], + "logprob_reward": -0.0008219677838496864, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008219677838496864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07592831552028656, + 0.1563195288181305, + 0.12747476994991302, + 0.056630443781614304, + 0.12479592114686966, + 0.11163539439439774, + 0.1175810918211937, + 0.05211323872208595, + 0.07103727012872696, + 0.10648400336503983 + ], + "credit_map": [ + 0.008231818675994873, + 0.00872915517538786, + 0.010097411461174488, + 0.0078808031976223, + 0.010212957859039307, + 0.009658502414822578, + 0.012252720072865486, + 0.008864951319992542, + 0.016585608944296837, + 0.016880057752132416, + 0.022797007113695145, + 0.013497548177838326, + 0.02803044766187668, + 0.022459669038653374, + 0.016736095771193504, + 0.01933310367166996, + 0.010046812705695629, + 0.013145051896572113, + 0.014789324253797531, + 0.010369202122092247, + 0.04182419925928116, + 0.010433461517095566, + 0.014634959399700165, + 0.012231756933033466, + 0.0074590761214494705, + 0.003320800606161356, + 0.012251903302967548, + 0.004766721278429031, + 0.005130475386977196, + 0.007776295766234398, + 0.0036533635575324297, + 0.012271813116967678, + 0.009148171171545982, + 0.015532457269728184, + 0.01645529270172119, + 0.023263681679964066, + 0.016991598531603813, + 0.011654572561383247, + 0.01861656829714775, + 0.01313359010964632, + 0.012379556894302368, + 0.009662079624831676, + 0.0186654981225729, + 0.019105641171336174, + 0.01172711793333292, + 0.013709204271435738, + 0.011618462391197681, + 0.014767838642001152, + 0.00892778392881155, + 0.011796325445175171, + 0.011904210783541203, + 0.02194502018392086, + 0.020142870023846626, + 0.007129636127501726, + 0.019764889031648636, + 0.015970364212989807, + 0.004943876061588526, + 0.0043859174475073814, + 0.0035093489568680525, + 0.011731080710887909, + 0.011371343396604061, + 0.0034879068844020367, + 0.005653258878737688, + 0.007030508480966091, + 0.006878705229610205, + 0.012984246015548706, + 0.008060368709266186, + 0.010307873599231243, + 0.01192031241953373, + 0.006468025967478752, + 0.007336082402616739, + 0.00708166416734457, + 0.015924938023090363, + 0.008511973544955254, + 0.007892491295933723, + 0.01184251718223095, + 0.021345948800444603, + 0.009139047004282475, + 0.01233311090618372, + 0.019493980333209038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.0008219677838496864 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04764551669359207, + 0.09144874662160873, + 0.10798927396535873, + 0.04019450768828392, + 0.39446038007736206, + 0.12053168565034866, + 0.05499390885233879, + 0.020466145128011703, + 0.06272746622562408, + 0.05954231694340706 + ], + "credit_map": [ + 0.004765976686030626, + 0.004754714202135801, + 0.006727562751621008, + 0.005341534037142992, + 0.005303174257278442, + 0.006967003457248211, + 0.005690319929271936, + 0.008095229975879192, + 0.008641590364277363, + 0.011352228932082653, + 0.012923145666718483, + 0.008267199620604515, + 0.018051737919449806, + 0.012017736211419106, + 0.010493614710867405, + 0.009701487608253956, + 0.008611095137894154, + 0.011860731057822704, + 0.012627876363694668, + 0.01011565700173378, + 0.033098701387643814, + 0.010174481198191643, + 0.011392598040401936, + 0.010108130984008312, + 0.005652406718581915, + 0.0037594656459987164, + 0.006072320509701967, + 0.004906074609607458, + 0.0045164404436945915, + 0.0037937976885586977, + 0.0065578073263168335, + 0.004936196841299534, + 0.022734763100743294, + 0.047261469066143036, + 0.04205930978059769, + 0.08416027575731277, + 0.04766470193862915, + 0.0275015439838171, + 0.08174756914377213, + 0.041330765932798386, + 0.012893104925751686, + 0.011549408547580242, + 0.019320368766784668, + 0.020025907084345818, + 0.013112843967974186, + 0.012666071765124798, + 0.015028848312795162, + 0.015935136005282402, + 0.003046714700758457, + 0.0066781602799892426, + 0.00606691837310791, + 0.014284948818385601, + 0.005757529754191637, + 0.003688979195430875, + 0.009375819936394691, + 0.006094838492572308, + 0.002421148121356964, + 0.0017139703268185258, + 0.0015867166221141815, + 0.0037872707471251488, + 0.0026631231885403395, + 0.0030843582935631275, + 0.0020461706444621086, + 0.003163388231769204, + 0.006284815259277821, + 0.010841315612196922, + 0.00770728150382638, + 0.008837537840008736, + 0.009392207488417625, + 0.006669271737337112, + 0.006662589963525534, + 0.0063324421644210815, + 0.011454325169324875, + 0.0048904563300311565, + 0.005976148881018162, + 0.006239103153347969, + 0.010420035570859909, + 0.007008849177509546, + 0.004601174965500832, + 0.00895222183316946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.00409490754827857 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035206615924835 + ], + "opposite_logprob_tokens": [ + -1.910352110862732 + ], + "logprob_reward": -0.16035206615924835, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035206615924835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11183209717273712, + 0.08515287190675735, + 0.108782559633255, + 0.06797236204147339, + 0.08154455572366714, + 0.09329143911600113, + 0.25096195936203003, + 0.0834747850894928, + 0.08961593359708786, + 0.02737145498394966 + ], + "credit_map": [ + 0.012210721150040627, + 0.00992101151496172, + 0.01602262258529663, + 0.01514859963208437, + 0.01565546542406082, + 0.011457116343080997, + 0.018198611214756966, + 0.013217947445809841, + 0.006919845938682556, + 0.013714232482016087, + 0.011792576871812344, + 0.01026371493935585, + 0.017352400347590446, + 0.008092844858765602, + 0.009779090993106365, + 0.0072381640784442425, + 0.009736491367220879, + 0.014222967438399792, + 0.01285775937139988, + 0.012613278813660145, + 0.022107519209384918, + 0.01075442973524332, + 0.015162934549152851, + 0.01132717914879322, + 0.004337329417467117, + 0.005902480334043503, + 0.009733449667692184, + 0.004212081432342529, + 0.023867236450314522, + 0.005189093295484781, + 0.00538764288648963, + 0.009343053214251995, + 0.0060626910999417305, + 0.011335054412484169, + 0.008620917797088623, + 0.015199677087366581, + 0.012806461192667484, + 0.004490087274461985, + 0.01563267409801483, + 0.0073969922959804535, + 0.010160956531763077, + 0.010604530572891235, + 0.012320016510784626, + 0.015048892237246037, + 0.013030525296926498, + 0.009942278265953064, + 0.01109333150088787, + 0.011090902611613274, + 0.014757640659809113, + 0.029426533728837967, + 0.024122053757309914, + 0.056781090795993805, + 0.03708459064364433, + 0.014256046153604984, + 0.043812576681375504, + 0.03072141483426094, + 0.008048939518630505, + 0.00629817321896553, + 0.008241179399192333, + 0.014165615662932396, + 0.01571538858115673, + 0.01007913425564766, + 0.009019953198730946, + 0.011906403116881847, + 0.007251131813973188, + 0.006193025503307581, + 0.02454734593629837, + 0.005975113715976477, + 0.008627098985016346, + 0.008207614533603191, + 0.006456241477280855, + 0.02235836163163185, + 0.0026540039107203484, + 0.0028382849413901567, + 0.0023877639323472977, + 0.0034114366862922907, + 0.004809219855815172, + 0.003542962484061718, + 0.0034297883976250887, + 0.004297995008528233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.16035206615924835 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2023443728685379, + 0.1344914734363556, + 0.09406761825084686, + 0.07083439826965332, + 0.12421289831399918, + 0.10295473784208298, + 0.11669305711984634, + 0.037188511341810226, + 0.0743338093161583, + 0.04287906363606453 + ], + "credit_map": [ + 0.024548178538680077, + 0.019669920206069946, + 0.03360694274306297, + 0.027642954140901566, + 0.01979215256869793, + 0.0223573986440897, + 0.028837092220783234, + 0.025889741256833076, + 0.011764013208448887, + 0.01693759486079216, + 0.023095013573765755, + 0.013693949207663536, + 0.019516395404934883, + 0.011907736770808697, + 0.014123777858912945, + 0.02345299907028675, + 0.006755566690117121, + 0.012514520436525345, + 0.012840802781283855, + 0.012983940541744232, + 0.024478573352098465, + 0.006544408388435841, + 0.010137636214494705, + 0.007812182884663343, + 0.00682327663525939, + 0.007153604179620743, + 0.014095719903707504, + 0.008232226595282555, + 0.007355425041168928, + 0.0072019933722913265, + 0.005838972982019186, + 0.014133177697658539, + 0.008437453769147396, + 0.01087129581719637, + 0.016817880794405937, + 0.02282928302884102, + 0.011851925402879715, + 0.014158527366816998, + 0.02575504407286644, + 0.013491492718458176, + 0.011094731278717518, + 0.01229880377650261, + 0.015111498534679413, + 0.01270765159279108, + 0.016966894268989563, + 0.010468422435224056, + 0.010366707108914852, + 0.013940028846263885, + 0.009193820878863335, + 0.010950304567813873, + 0.011676985770463943, + 0.0235426165163517, + 0.018363170325756073, + 0.009956014342606068, + 0.01715102791786194, + 0.01585911028087139, + 0.004308219067752361, + 0.0032543963752686977, + 0.004310106858611107, + 0.005271241534501314, + 0.0066322688944637775, + 0.004870236851274967, + 0.004371222574263811, + 0.0041708205826580524, + 0.00844461377710104, + 0.009743301197886467, + 0.01208366360515356, + 0.008814929984509945, + 0.011278296820819378, + 0.007042249199002981, + 0.00868032407015562, + 0.008246433921158314, + 0.006475798320025206, + 0.0025987245608121157, + 0.004276866093277931, + 0.004995544441044331, + 0.005975414998829365, + 0.0064516132697463036, + 0.005708651151508093, + 0.006396451499313116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005256047938019037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tilted Dutch-angle wide-angle lens, with an unsteady tracking motion, captures a surreal stroll through a city late at night. A massive, bioluminescent siphonophore, the size of a parade float, floats slowly and majestically in the center of a deserted street. Its translucent, gelatinous body, composed of countless tiny, interconnected individuals, emanates a cold blue and purple glow, a stark contrast to the warm, flickering neon lights of its surroundings. Its massive, umbrella-shaped propulsion organs rhythmically contract and expand, each beat pulsing outward like a heart, a brilliant cyan halo and accompanied by a clear, ethereal cry that echoes between the towering skyscrapers, whose windows are mostly darkened. The damp asphalt reflects the colorful lights of the signs and the luminescent glow of the strange creature itself like a mirror. Countless luminescent tentacles dangle from its lower body like ornate tassels, swaying gently, occasionally skimming the ground, leaving fleeting traces of phosphorescence on the damp pavement. The entire picture appears both dreamy and unsettling due to the tilted composition, creating a grand, lonely and awe-inspiring visual spectacle.", + "seed": 31853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_15/index_8592_seed_31853_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": -0.007935524580534547, + "vlm_logprob": -0.007935524580534547, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.007935524580534547, + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013581139966845512 + ], + "opposite_logprob_tokens": [ + -4.388581275939941 + ], + "logprob_reward": -0.013581139966845512, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013581139966845512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13312159478664398, + 0.06215599179267883, + 0.02348335273563862, + 0.04342057928442955, + 0.11507399380207062, + 0.15375317633152008, + 0.044087767601013184, + 0.26312509179115295, + 0.08268994092941284, + 0.07908856123685837 + ], + "credit_map": [ + 0.015656528994441032, + 0.022235043346881866, + 0.016108375042676926, + 0.01703120768070221, + 0.016553951427340508, + 0.01187765970826149, + 0.014951487071812153, + 0.018707331269979477, + 0.005881934426724911, + 0.005072812084108591, + 0.00686196144670248, + 0.006345496512949467, + 0.008974041789770126, + 0.013173321262001991, + 0.006790950428694487, + 0.009055474773049355, + 0.001904950593598187, + 0.0036643771454691887, + 0.003319191513583064, + 0.0026279371231794357, + 0.003318999893963337, + 0.0033660917542874813, + 0.0024121468886733055, + 0.002869656542316079, + 0.0038738229777663946, + 0.00573120079934597, + 0.006002491340041161, + 0.004592164885252714, + 0.00714291725307703, + 0.005032294429838657, + 0.00566935446113348, + 0.005376332439482212, + 0.013951282948255539, + 0.015482114627957344, + 0.015087245963513851, + 0.01561069954186678, + 0.01570119708776474, + 0.014567938633263111, + 0.013431002385914326, + 0.011242508888244629, + 0.009017851203680038, + 0.013449875637888908, + 0.0296791959553957, + 0.014478731900453568, + 0.03980392962694168, + 0.009796427562832832, + 0.03184615448117256, + 0.005681018810719252, + 0.0055345878936350346, + 0.0051629748195409775, + 0.007578195072710514, + 0.0055922893807291985, + 0.003914997447282076, + 0.005103015340864658, + 0.0046168128028512, + 0.006584892515093088, + 0.019915299490094185, + 0.054168570786714554, + 0.026752835139632225, + 0.040542326867580414, + 0.056637413799762726, + 0.017465095967054367, + 0.026734977960586548, + 0.020908568054437637, + 0.009295104071497917, + 0.006377631798386574, + 0.022616544738411903, + 0.011239778250455856, + 0.006972678937017918, + 0.007814615033566952, + 0.008237062953412533, + 0.010136528871953487, + 0.007692462299019098, + 0.006771608255803585, + 0.016667477786540985, + 0.010655513033270836, + 0.007378775160759687, + 0.008613432757556438, + 0.00927579216659069, + 0.012033491395413876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.013581139966845512 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007238152320496738 + ], + "opposite_logprob_tokens": [ + -7.250723838806152 + ], + "logprob_reward": -0.0007238152320496738, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007238152320496738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14249052107334137, + 0.11254392564296722, + 0.07133828848600388, + 0.06778751313686371, + 0.14296218752861023, + 0.09474728256464005, + 0.08519253879785538, + 0.12128927558660507, + 0.03592829778790474, + 0.12572011351585388 + ], + "credit_map": [ + 0.01473784726113081, + 0.011298729106783867, + 0.019511761143803596, + 0.01691655069589615, + 0.019276088103652, + 0.014551185071468353, + 0.01949562318623066, + 0.02670273557305336, + 0.008007279597222805, + 0.024203574284911156, + 0.014471909962594509, + 0.009579298086464405, + 0.02103768102824688, + 0.0114649822935462, + 0.012424907647073269, + 0.011354291811585426, + 0.007232696749269962, + 0.011139914393424988, + 0.011894484050571918, + 0.007372685242444277, + 0.009025363251566887, + 0.008746758103370667, + 0.006000495981425047, + 0.009925897233188152, + 0.00907757505774498, + 0.00888748001307249, + 0.010872086510062218, + 0.0049674417823553085, + 0.009100900031626225, + 0.007030000910162926, + 0.006481655407696962, + 0.011370382271707058, + 0.017474083229899406, + 0.020912645384669304, + 0.019149357452988625, + 0.02128428779542446, + 0.019614674150943756, + 0.01458668988198042, + 0.01783020608127117, + 0.012110252864658833, + 0.00627308851107955, + 0.009392527863383293, + 0.018454715609550476, + 0.009749525226652622, + 0.024078475311398506, + 0.006674152333289385, + 0.016281500458717346, + 0.0038432928267866373, + 0.006516775581985712, + 0.010037417523562908, + 0.011008224450051785, + 0.010918112471699715, + 0.011678212322294712, + 0.008489953354001045, + 0.018900994211435318, + 0.007642847951501608, + 0.008948939852416515, + 0.02329319156706333, + 0.012500828132033348, + 0.0215840432792902, + 0.02539929188787937, + 0.008373681455850601, + 0.012219248339533806, + 0.008970052003860474, + 0.003679007524624467, + 0.003369775367900729, + 0.00451057031750679, + 0.006290016230195761, + 0.0024424041621387005, + 0.004537729546427727, + 0.006112990900874138, + 0.0049858056008815765, + 0.011887582018971443, + 0.016561875119805336, + 0.020150456577539444, + 0.016303852200508118, + 0.014470375142991543, + 0.018961306661367416, + 0.011620967648923397, + 0.01576370932161808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0007238152320496738 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615867579355836 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.003615867579355836, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615867579355836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06918249279260635, + 0.09338594973087311, + 0.05608608201146126, + 0.07294318825006485, + 0.05978138744831085, + 0.400817334651947, + 0.0357101745903492, + 0.03892422839999199, + 0.07217118889093399, + 0.10099798440933228 + ], + "credit_map": [ + 0.012083900161087513, + 0.006656324956566095, + 0.007485473994165659, + 0.009525997564196587, + 0.009242082946002483, + 0.007351238746196032, + 0.009940453805029392, + 0.006897021550685167, + 0.009886959567666054, + 0.008309938944876194, + 0.014383179135620594, + 0.010570734739303589, + 0.013571127317845821, + 0.01301766186952591, + 0.010134832933545113, + 0.013511527329683304, + 0.0037413130048662424, + 0.00939978938549757, + 0.011379029601812363, + 0.005268022883683443, + 0.006882523186504841, + 0.005844514816999435, + 0.00472335796803236, + 0.008847531862556934, + 0.006495234090834856, + 0.009448089636862278, + 0.010725309140980244, + 0.006983927916735411, + 0.008334002457559109, + 0.009671895764768124, + 0.011825709603726864, + 0.00945901870727539, + 0.007979167625308037, + 0.007128095254302025, + 0.007819083519279957, + 0.006490449886769056, + 0.007344404701143503, + 0.007132244762033224, + 0.00912844855338335, + 0.00675949128344655, + 0.021648721769452095, + 0.03591950237751007, + 0.06403667479753494, + 0.0400947742164135, + 0.0764031857252121, + 0.03595351800322533, + 0.10342424362897873, + 0.02333669736981392, + 0.0033538653515279293, + 0.004533383995294571, + 0.006215515546500683, + 0.004580676089972258, + 0.005453481338918209, + 0.0034918601159006357, + 0.0048212516121566296, + 0.003260140074416995, + 0.0033134089317172766, + 0.007267369888722897, + 0.004435062874108553, + 0.0045348783023655415, + 0.0077525172382593155, + 0.0030343651305884123, + 0.004054976627230644, + 0.004531646613031626, + 0.0074066054075956345, + 0.005849577020853758, + 0.01627529226243496, + 0.008809980936348438, + 0.007376582827419043, + 0.007984260097146034, + 0.009187745861709118, + 0.009281150996685028, + 0.010491850785911083, + 0.014794845134019852, + 0.013311001472175121, + 0.01040616910904646, + 0.01542974915355444, + 0.013214819133281708, + 0.011694471351802349, + 0.01165507361292839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.003615867579355836 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575299859046936 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575299859046936, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575299859046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08853589743375778, + 0.0409356914460659, + 0.06452431529760361, + 0.08977010846138, + 0.058534879237413406, + 0.05920366570353508, + 0.22527901828289032, + 0.31071770191192627, + 0.029012074694037437, + 0.033486757427453995 + ], + "credit_map": [ + 0.009571653790771961, + 0.007028740830719471, + 0.019420647993683815, + 0.008502637967467308, + 0.0069624753668904305, + 0.007617977913469076, + 0.006451993715018034, + 0.022979754954576492, + 0.004641341511160135, + 0.004298144020140171, + 0.003651607083156705, + 0.005430967081338167, + 0.006956335622817278, + 0.006539409514516592, + 0.004887607879936695, + 0.00453027430921793, + 0.005404983647167683, + 0.004350487608462572, + 0.010184461250901222, + 0.007039542309939861, + 0.008311505429446697, + 0.011067287996411324, + 0.007845111191272736, + 0.010320926085114479, + 0.0065491655841469765, + 0.010774102993309498, + 0.011949737556278706, + 0.009329844266176224, + 0.011184037663042545, + 0.009311691857874393, + 0.01886315643787384, + 0.011808361858129501, + 0.010448714718222618, + 0.00592344393953681, + 0.008285640738904476, + 0.006785888224840164, + 0.005697471089661121, + 0.007621621247380972, + 0.0057166749611496925, + 0.00805541593581438, + 0.004192663822323084, + 0.005796033889055252, + 0.009139061905443668, + 0.008038447238504887, + 0.008697768673300743, + 0.004256340209394693, + 0.013462856411933899, + 0.005620486102998257, + 0.01653355546295643, + 0.028749000281095505, + 0.02928204834461212, + 0.02980024553835392, + 0.03343255817890167, + 0.015905817970633507, + 0.05194414034485817, + 0.01963162049651146, + 0.025086266919970512, + 0.07518677413463593, + 0.02465924620628357, + 0.04364967718720436, + 0.07065965980291367, + 0.018946999683976173, + 0.03188106417655945, + 0.020647983998060226, + 0.0034546698443591595, + 0.0024929638020694256, + 0.002989049768075347, + 0.006391232833266258, + 0.0023893381003290415, + 0.0025989003479480743, + 0.0033330663572996855, + 0.005362850148230791, + 0.005047206301242113, + 0.002758678514510393, + 0.006900651380419731, + 0.004383104853332043, + 0.0028461019974201918, + 0.003937204368412495, + 0.002373200375586748, + 0.005240604747086763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.020575299859046936 + }, + { + "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 bioluminescent siphonophore floating in the center of a deserted city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the creature have a translucent gelatinous body with a cold blue and purple glow and umbrella-shaped propulsion organs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are countless luminescent tentacles visible dangling from the creature's lower body and swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a tilted Dutch-angle wide-angle view with unsteady tracking motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the damp asphalt reflect the colorful neon lights and the creature's luminescent glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011815002653747797 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011815002653747797, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011815002653747797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12443914264440536, + 0.10929291695356369, + 0.07067606598138809, + 0.03845047578215599, + 0.047378625720739365, + 0.20701812207698822, + 0.04613492637872696, + 0.10558826476335526, + 0.15986332297325134, + 0.09115812927484512 + ], + "credit_map": [ + 0.015465699136257172, + 0.012892555445432663, + 0.018279001116752625, + 0.01840830221772194, + 0.011418469250202179, + 0.010737271048128605, + 0.021773450076580048, + 0.015464394353330135, + 0.008682244457304478, + 0.01769898645579815, + 0.017144713550806046, + 0.010739651508629322, + 0.017568478360772133, + 0.011616876348853111, + 0.011361385695636272, + 0.014480579644441605, + 0.006892036180943251, + 0.012829395942389965, + 0.008569255471229553, + 0.007748522330075502, + 0.010896530002355576, + 0.008195501752197742, + 0.006696755066514015, + 0.008848069235682487, + 0.002737711649388075, + 0.0045713623985648155, + 0.004555721301585436, + 0.005198871251195669, + 0.005058695562183857, + 0.004710309207439423, + 0.006582801230251789, + 0.005035002250224352, + 0.004789297003298998, + 0.006233530584722757, + 0.006477777846157551, + 0.006683203857392073, + 0.0068474868312478065, + 0.00693862047046423, + 0.004264362622052431, + 0.0051443446427583694, + 0.01549201924353838, + 0.01578558050096035, + 0.05049079284071922, + 0.018789639696478844, + 0.06567490100860596, + 0.015421019867062569, + 0.017026694491505623, + 0.008337465114891529, + 0.005478880368173122, + 0.004174289759248495, + 0.006845848169177771, + 0.005790500435978174, + 0.005062242969870567, + 0.0051300739869475365, + 0.006920629180967808, + 0.006732459180057049, + 0.01032426580786705, + 0.01463597733527422, + 0.01804363913834095, + 0.01602938026189804, + 0.0177113376557827, + 0.00834348052740097, + 0.01078085508197546, + 0.009719334542751312, + 0.017493724822998047, + 0.012324885465204716, + 0.03764498978853226, + 0.018794473260641098, + 0.015623845160007477, + 0.018380949273705482, + 0.01967155747115612, + 0.019928893074393272, + 0.013188390992581844, + 0.006544045172631741, + 0.01598253659904003, + 0.01396337989717722, + 0.007242585998028517, + 0.012319318950176239, + 0.007885153405368328, + 0.014032714068889618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0011815002653747797 + } + ], + "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/0000167/group_2/index_tensor([3206, 3206])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_tensor([3206, 3206])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e190d75002a9f6cf73ef1df53a637af2b70325a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_tensor([3206, 3206])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly tilted-up close-up captures the viewer's gaze firmly on the roar of a white tiger. Dramatic, high-contrast lighting cast from the side and above, sculpting every detail with pinpoint precision. The tiger's thick, snow-like fur is crisscrossed by inky black stripes that snake along its powerful facial features, strained with intense rage. Its mouth is stretched to its limit, a terrifying maw revealing a cavernous, dark pink interior. A massive, barbed tongue curls back in its roar, while ivory canines several inches long, dagger-sharp, ivory-colored canines gleam with a wet, cold gleam. A deafening roar erupts from deep within its throat, its wet black nose wrinkling violently in aggression, its thick white beard quivering with the force, and tiny drops of saliva, driven by the mighty sound, form a fleeting mist in the air. The background is completely blurred into an abstract block of dark green and brown, placing the undisputed focus on this face, full of primal power and endless ferocity, creating a terrifying and tense atmosphere where one can almost feel the air vibrate.", + "seed": 19642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19642_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": -0.009172077942639589, + "vlm_logprob": -0.009172077942639589, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.009172077942639589, + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006450782995671034 + ], + "opposite_logprob_tokens": [ + -6.756450653076172 + ], + "logprob_reward": -0.006450782995671034, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.006450782995671034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09753657877445221, + 0.07587110996246338, + 0.18800899386405945, + 0.08254736661911011, + 0.2323971837759018, + 0.03492411971092224, + 0.09056495130062103, + 0.11508502811193466, + 0.03757018968462944, + 0.045494548976421356 + ], + "credit_map": [ + 0.0101446732878685, + 0.008885369636118412, + 0.01655099168419838, + 0.010237975046038628, + 0.010035772807896137, + 0.01376622449606657, + 0.015963872894644737, + 0.011951700784265995, + 0.007518807891756296, + 0.006952873896807432, + 0.019141284748911858, + 0.014181687496602535, + 0.005569247528910637, + 0.006736461538821459, + 0.0065351794473826885, + 0.009235566481947899, + 0.03688492998480797, + 0.018266471102833748, + 0.06266438215970993, + 0.020246470347046852, + 0.010404989123344421, + 0.013431088998913765, + 0.009962190873920918, + 0.016148477792739868, + 0.010359260253608227, + 0.005939907394349575, + 0.009146860800683498, + 0.005836538504809141, + 0.01848471723496914, + 0.009069289080798626, + 0.012617426924407482, + 0.01109336968511343, + 0.023790117353200912, + 0.021603664383292198, + 0.06704077869653702, + 0.028478754684329033, + 0.016322053968906403, + 0.019212817773222923, + 0.022634103894233704, + 0.03331487625837326, + 0.004405419807881117, + 0.003571880981326103, + 0.00693919975310564, + 0.004284687805920839, + 0.00348356063477695, + 0.0030207897070795298, + 0.004943337757140398, + 0.004275243729352951, + 0.009224099107086658, + 0.012730713002383709, + 0.010242610238492489, + 0.014389085583388805, + 0.011084609664976597, + 0.015364806167781353, + 0.008227640762925148, + 0.009301380254328251, + 0.012596812099218369, + 0.01249796710908413, + 0.017329631373286247, + 0.013662018813192844, + 0.015493866056203842, + 0.011622324585914612, + 0.018594687804579735, + 0.013287723995745182, + 0.0031051095575094223, + 0.0035068257711827755, + 0.0054305680096149445, + 0.005905665922909975, + 0.004368284251540899, + 0.00555112911388278, + 0.005106410011649132, + 0.00459619564935565, + 0.002931082621216774, + 0.006388519890606403, + 0.008572575636208057, + 0.004572638776153326, + 0.0048735737800598145, + 0.006291280034929514, + 0.005624890327453613, + 0.006239987909793854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.006450782995671034 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032014804892241955 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.0032014804892241955, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032014804892241955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052489642053842545, + 0.061450473964214325, + 0.08031143248081207, + 0.22760654985904694, + 0.050650905817747116, + 0.08268268406391144, + 0.07427040487527847, + 0.042857591062784195, + 0.2937595844268799, + 0.03392074629664421 + ], + "credit_map": [ + 0.005069687031209469, + 0.007002883590757847, + 0.006282642018049955, + 0.006120228208601475, + 0.005696193780750036, + 0.007577711716294289, + 0.006659162696450949, + 0.008081133477389812, + 0.005560608580708504, + 0.009281744249165058, + 0.00558013329282403, + 0.012968774884939194, + 0.008876897394657135, + 0.007265911437571049, + 0.006059804931282997, + 0.005856597330421209, + 0.016125096008181572, + 0.007358719129115343, + 0.020562369376420975, + 0.009544914588332176, + 0.005626622587442398, + 0.0056847454980015755, + 0.006030350457876921, + 0.009378613904118538, + 0.018786635249853134, + 0.011598552577197552, + 0.025344694033265114, + 0.017951957881450653, + 0.0238060113042593, + 0.025207173079252243, + 0.0602453388273716, + 0.04466620087623596, + 0.004911435768008232, + 0.006185783073306084, + 0.010333920828998089, + 0.0054780482314527035, + 0.005110799800604582, + 0.005718403495848179, + 0.005779987201094627, + 0.00713252741843462, + 0.00991110410541296, + 0.009196735918521881, + 0.013893356546759605, + 0.008196908049285412, + 0.007406860124319792, + 0.011263345368206501, + 0.013582088984549046, + 0.009232282638549805, + 0.007305544335395098, + 0.009044546633958817, + 0.006777995266020298, + 0.01613541878759861, + 0.009262175299227238, + 0.006922503001987934, + 0.008358759805560112, + 0.010463462211191654, + 0.004850289318710566, + 0.00697682099416852, + 0.006894739344716072, + 0.004247149918228388, + 0.005189287476241589, + 0.005102427210658789, + 0.005063964985311031, + 0.00453291367739439, + 0.01945056952536106, + 0.01815439574420452, + 0.07008154690265656, + 0.043286532163619995, + 0.019225407391786575, + 0.03764500096440315, + 0.04253227263689041, + 0.043383851647377014, + 0.0023542039562016726, + 0.0035622783470898867, + 0.005372995510697365, + 0.0028365610633045435, + 0.004570815712213516, + 0.004643300548195839, + 0.006448676344007254, + 0.004131916910409927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.0032014804892241955 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04631247743964195, + 0.03751863166689873, + 0.057309336960315704, + 0.047356270253658295, + 0.1524038314819336, + 0.03273522108793259, + 0.16667890548706055, + 0.03711305931210518, + 0.4037124812602997, + 0.018859738484025 + ], + "credit_map": [ + 0.004490553867071867, + 0.00781956035643816, + 0.00762338237836957, + 0.004525813274085522, + 0.004197350703179836, + 0.004379820078611374, + 0.005545506719499826, + 0.0077304900623857975, + 0.004051426891237497, + 0.0038024326786398888, + 0.0055504534393548965, + 0.006677537225186825, + 0.004759566858410835, + 0.003585435915738344, + 0.004299203399568796, + 0.0047925738617777824, + 0.004199442453682423, + 0.0072451066225767136, + 0.012942757457494736, + 0.009927261620759964, + 0.006074828561395407, + 0.0049753435887396336, + 0.00595693988725543, + 0.005987655837088823, + 0.005307060666382313, + 0.004491494502872229, + 0.004623511340469122, + 0.003929058089852333, + 0.009477073326706886, + 0.003932361025363207, + 0.010257666930556297, + 0.005338048096746206, + 0.010702426545321941, + 0.009819789789617062, + 0.05162171274423599, + 0.011923380196094513, + 0.007235128898173571, + 0.016452142968773842, + 0.013527930714190006, + 0.03112131915986538, + 0.0037544139195233583, + 0.0031916603911668062, + 0.003322840901091695, + 0.00516148004680872, + 0.0037968887481838465, + 0.004782072268426418, + 0.005145842209458351, + 0.0035800246987491846, + 0.017517494037747383, + 0.017358895391225815, + 0.018174413591623306, + 0.0395381785929203, + 0.018671488389372826, + 0.01698055863380432, + 0.015245743095874786, + 0.023192135617136955, + 0.0039023265708237886, + 0.004045387730002403, + 0.005482351407408714, + 0.003746203612536192, + 0.0041779945604503155, + 0.004807302728295326, + 0.0068407319486141205, + 0.004110760986804962, + 0.02642064355313778, + 0.026307346299290657, + 0.09540553390979767, + 0.06237129122018814, + 0.024318767711520195, + 0.05630524829030037, + 0.056054506450891495, + 0.05652914568781853, + 0.0023266756907105446, + 0.0021498703863471746, + 0.002719808602705598, + 0.0014660439919680357, + 0.0022888188250362873, + 0.00394192012026906, + 0.0018399989930912852, + 0.0021266015246510506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.005946444813162088 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19594396650791168, + 0.07408774644136429, + 0.16382116079330444, + 0.09864851832389832, + 0.1262112408876419, + 0.04776889830827713, + 0.15391366183757782, + 0.0580546073615551, + 0.04740947112441063, + 0.034140702337026596 + ], + "credit_map": [ + 0.02429812401533127, + 0.04042675346136093, + 0.019674966111779213, + 0.03430362790822983, + 0.01670360192656517, + 0.018622826784849167, + 0.020593706518411636, + 0.021320348605513573, + 0.008473529480397701, + 0.006559547036886215, + 0.007632218301296234, + 0.014420631341636181, + 0.009903743863105774, + 0.01115928404033184, + 0.006497363559901714, + 0.009441433474421501, + 0.014383028261363506, + 0.015802770853042603, + 0.0419406034052372, + 0.027016589418053627, + 0.01210102904587984, + 0.015778815373778343, + 0.01579062081873417, + 0.02100769430398941, + 0.011295150965452194, + 0.005970103666186333, + 0.010278950445353985, + 0.00695928605273366, + 0.017652662470936775, + 0.008617081679403782, + 0.018525345250964165, + 0.019349928945302963, + 0.016230016946792603, + 0.016746703535318375, + 0.02491096407175064, + 0.01659688912332058, + 0.011613472364842892, + 0.01578465849161148, + 0.010456454008817673, + 0.013872078619897366, + 0.005034165922552347, + 0.003674105741083622, + 0.00719141960144043, + 0.005308950785547495, + 0.006153232418000698, + 0.006944998167455196, + 0.0074043869972229, + 0.0060576386749744415, + 0.015432323329150677, + 0.013037815690040588, + 0.012009559199213982, + 0.036971062421798706, + 0.011217731982469559, + 0.025505928322672844, + 0.011349746957421303, + 0.028389494866132736, + 0.006539317313581705, + 0.0060448357835412025, + 0.009600122459232807, + 0.007326685357838869, + 0.007202936336398125, + 0.005446874536573887, + 0.00945152435451746, + 0.00644230842590332, + 0.003908785991370678, + 0.004833359271287918, + 0.010257837362587452, + 0.006741080433130264, + 0.003948014695197344, + 0.007090514991432428, + 0.004633951932191849, + 0.005995927378535271, + 0.0031091340351849794, + 0.004723673686385155, + 0.003755728481337428, + 0.0034207950811833143, + 0.004288023803383112, + 0.0069886124692857265, + 0.003277338808402419, + 0.004577395040541887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.01419964898377657 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606203243136406 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606203243136406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606203243136406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0927705466747284, + 0.11198040843009949, + 0.08685066550970078, + 0.18280398845672607, + 0.28396764397621155, + 0.011570094153285027, + 0.059100110083818436, + 0.02523697167634964, + 0.036730095744132996, + 0.10898945480585098 + ], + "credit_map": [ + 0.009686946868896484, + 0.011346789076924324, + 0.0126658845692873, + 0.00894674751907587, + 0.010312039405107498, + 0.017860673367977142, + 0.01123744621872902, + 0.01071402058005333, + 0.01136435940861702, + 0.008788256905972958, + 0.011256138794124126, + 0.01978152245283127, + 0.01830848678946495, + 0.01562329288572073, + 0.01227596029639244, + 0.014582392759621143, + 0.019454441964626312, + 0.004839129280298948, + 0.027525829151272774, + 0.009887310676276684, + 0.004250324796885252, + 0.006573760416358709, + 0.005517338868230581, + 0.008802535012364388, + 0.013376763090491295, + 0.007651010528206825, + 0.01700025051832199, + 0.011605183593928814, + 0.022893426939845085, + 0.017207682132720947, + 0.04190110042691231, + 0.051168572157621384, + 0.027894839644432068, + 0.032073963433504105, + 0.07767642289400101, + 0.030880849808454514, + 0.019835328683257103, + 0.030638059601187706, + 0.023662416264414787, + 0.0413057915866375, + 0.0014861341333016753, + 0.0017906541470438242, + 0.0014517587842419744, + 0.001582099124789238, + 0.0011710267281159759, + 0.0012266812846064568, + 0.0017142490250989795, + 0.0011474910425022244, + 0.0063646226190030575, + 0.007105794735252857, + 0.0052278428338468075, + 0.010668440721929073, + 0.006977117620408535, + 0.007237328216433525, + 0.007505648769438267, + 0.00801330991089344, + 0.0034097193274646997, + 0.0040635401383042336, + 0.0030214055441319942, + 0.0027589944656938314, + 0.0027063139714300632, + 0.0030489189084619284, + 0.0029586637392640114, + 0.0032694158144295216, + 0.002136198105290532, + 0.003322007367387414, + 0.005990323144942522, + 0.0055188764818012714, + 0.001988181844353676, + 0.006500863470137119, + 0.0038285679183900356, + 0.007445076946169138, + 0.009064896032214165, + 0.01820978708565235, + 0.014083903282880783, + 0.014009648934006691, + 0.013933056965470314, + 0.013786766678094864, + 0.013997816480696201, + 0.011903582140803337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.01606203243136406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly tilted-up close-up captures the viewer's gaze firmly on the roar of a white tiger. Dramatic, high-contrast lighting cast from the side and above, sculpting every detail with pinpoint precision. The tiger's thick, snow-like fur is crisscrossed by inky black stripes that snake along its powerful facial features, strained with intense rage. Its mouth is stretched to its limit, a terrifying maw revealing a cavernous, dark pink interior. A massive, barbed tongue curls back in its roar, while ivory canines several inches long, dagger-sharp, ivory-colored canines gleam with a wet, cold gleam. A deafening roar erupts from deep within its throat, its wet black nose wrinkling violently in aggression, its thick white beard quivering with the force, and tiny drops of saliva, driven by the mighty sound, form a fleeting mist in the air. The background is completely blurred into an abstract block of dark green and brown, placing the undisputed focus on this face, full of primal power and endless ferocity, creating a terrifying and tense atmosphere where one can almost feel the air vibrate.", + "seed": 19643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19643_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": -0.00949025652371347, + "vlm_logprob": -0.00949025652371347, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.00949025652371347, + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010380445048213005 + ], + "opposite_logprob_tokens": [ + -6.385380268096924 + ], + "logprob_reward": -0.010380445048213005, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.010380445048213005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30465075373649597, + 0.1021164134144783, + 0.10101327300071716, + 0.08618921786546707, + 0.1082189530134201, + 0.05356314033269882, + 0.09444849193096161, + 0.041169777512550354, + 0.03459014743566513, + 0.07403983920812607 + ], + "credit_map": [ + 0.028319550678133965, + 0.033297497779130936, + 0.05060933530330658, + 0.03971376270055771, + 0.044033415615558624, + 0.039167944341897964, + 0.036929626017808914, + 0.03257961571216583, + 0.009225030429661274, + 0.011104428209364414, + 0.01213087234646082, + 0.016068235039711, + 0.009423592127859592, + 0.02121810056269169, + 0.010319644585251808, + 0.012626513838768005, + 0.015693487599492073, + 0.008182584308087826, + 0.013212759047746658, + 0.01273233164101839, + 0.01123861875385046, + 0.013373276218771935, + 0.011749591678380966, + 0.014830621890723705, + 0.012260094285011292, + 0.010549700818955898, + 0.01371676754206419, + 0.00965036265552044, + 0.008203363977372646, + 0.012407638132572174, + 0.007908200845122337, + 0.011493097990751266, + 0.009990575723350048, + 0.011009707115590572, + 0.012173889204859734, + 0.015883658081293106, + 0.014192230068147182, + 0.014113472774624825, + 0.017242934554815292, + 0.013612481765449047, + 0.0059036933816969395, + 0.004707994405180216, + 0.0054616923443973064, + 0.007978743873536587, + 0.005825265310704708, + 0.009378457441926003, + 0.006327528972178698, + 0.007979764603078365, + 0.017905956134200096, + 0.006304959300905466, + 0.01709059439599514, + 0.010918653570115566, + 0.006804066710174084, + 0.007890728302299976, + 0.014061322435736656, + 0.013472218066453934, + 0.004703710786998272, + 0.005152921192348003, + 0.005416559055447578, + 0.008524934761226177, + 0.0044613485224545, + 0.003713647834956646, + 0.005494055338203907, + 0.0037026016507297754, + 0.003929643891751766, + 0.00432704109698534, + 0.006276941858232021, + 0.005918869748711586, + 0.004344604443758726, + 0.002535915235057473, + 0.0037790194619446993, + 0.003478110069409013, + 0.006519781891256571, + 0.00620135385543108, + 0.008573098108172417, + 0.020566541701555252, + 0.008052535355091095, + 0.005637789145112038, + 0.0068616485223174095, + 0.011627084575593472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.010380445048213005 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002506331307813525 + ], + "opposite_logprob_tokens": [ + -6.002506256103516 + ], + "logprob_reward": -0.002506331307813525, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002506331307813525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17615823447704315, + 0.15906664729118347, + 0.11106216907501221, + 0.07628235965967178, + 0.12802129983901978, + 0.06767111271619797, + 0.0673782080411911, + 0.07950841635465622, + 0.055023230612277985, + 0.07982838898897171 + ], + "credit_map": [ + 0.01441065315157175, + 0.014398898929357529, + 0.025411516427993774, + 0.025482848286628723, + 0.020715191960334778, + 0.0231898445636034, + 0.022189481183886528, + 0.030359797179698944, + 0.013703477568924427, + 0.02061353623867035, + 0.01688443310558796, + 0.027247359976172447, + 0.014663776382803917, + 0.020858068019151688, + 0.024229727685451508, + 0.02086627669632435, + 0.012347354553639889, + 0.010795251466333866, + 0.015165679156780243, + 0.013220390304923058, + 0.013489467091858387, + 0.01713290810585022, + 0.014127865433692932, + 0.01478324644267559, + 0.0049094464629888535, + 0.008473066613078117, + 0.010945106856524944, + 0.009256836026906967, + 0.010395356453955173, + 0.010657420381903648, + 0.012340881861746311, + 0.009304244071245193, + 0.009072571992874146, + 0.012732496485114098, + 0.017880957573652267, + 0.01953125186264515, + 0.015587969683110714, + 0.01391158439218998, + 0.023523375391960144, + 0.01578107848763466, + 0.012507491745054722, + 0.005939969327300787, + 0.007557533215731382, + 0.010324167087674141, + 0.008048716932535172, + 0.008617724291980267, + 0.006622752640396357, + 0.008052753284573555, + 0.007903732359409332, + 0.004822871182113886, + 0.0123788733035326, + 0.013118763454258442, + 0.004768174607306719, + 0.004865818656980991, + 0.008428385481238365, + 0.011091583408415318, + 0.00955197960138321, + 0.007111699786037207, + 0.009008967317640781, + 0.01930544152855873, + 0.0073266117833554745, + 0.011100715957581997, + 0.00783927459269762, + 0.008263724856078625, + 0.005903872661292553, + 0.009080933406949043, + 0.008632637560367584, + 0.010324371978640556, + 0.005783163011074066, + 0.004706831648945808, + 0.004150996450334787, + 0.006440426222980022, + 0.007762487046420574, + 0.007588572334498167, + 0.010529940016567707, + 0.02117445506155491, + 0.005681383423507214, + 0.007522163912653923, + 0.0067131961695849895, + 0.01285618543624878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002506331307813525 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539827264845371 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539827264845371, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539827264845371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1601865440607071, + 0.06276034563779831, + 0.10488123446702957, + 0.07867611199617386, + 0.20817439258098602, + 0.03363805636763573, + 0.13985121250152588, + 0.07303760200738907, + 0.03812588006258011, + 0.10066864639520645 + ], + "credit_map": [ + 0.015572811476886272, + 0.012522452510893345, + 0.02165965735912323, + 0.022651830688118935, + 0.01971631683409214, + 0.021009529009461403, + 0.019904376938939095, + 0.027149561792612076, + 0.006990955676883459, + 0.0069159106351435184, + 0.006134785711765289, + 0.010006150230765343, + 0.006032396573573351, + 0.010935239493846893, + 0.008611636236310005, + 0.007133267819881439, + 0.010374428704380989, + 0.008071556687355042, + 0.014520169235765934, + 0.012992048636078835, + 0.010738905519247055, + 0.018654389306902885, + 0.013962001539766788, + 0.015567734837532043, + 0.009232875891029835, + 0.008345584385097027, + 0.01095129270106554, + 0.012171118520200253, + 0.009248948656022549, + 0.010635619051754475, + 0.007997710257768631, + 0.0100929606705904, + 0.014910588040947914, + 0.02684360183775425, + 0.021016962826251984, + 0.03136886656284332, + 0.023562638089060783, + 0.030525173991918564, + 0.03380206599831581, + 0.0261444840580225, + 0.004169447813183069, + 0.003566617611795664, + 0.004685052204877138, + 0.004465300589799881, + 0.004003929439932108, + 0.0035190905909985304, + 0.004793951753526926, + 0.004434668458998203, + 0.01597314141690731, + 0.01026079524308443, + 0.029130976647138596, + 0.02261575683951378, + 0.010961491614580154, + 0.009950067847967148, + 0.01779930666089058, + 0.023159680888056755, + 0.00824085995554924, + 0.005286396946758032, + 0.007089004386216402, + 0.010925652459263802, + 0.008613322861492634, + 0.011214206926524639, + 0.014477279968559742, + 0.007190880831331015, + 0.004158738534897566, + 0.0054641240276396275, + 0.004914767108857632, + 0.004738600924611092, + 0.004548335913568735, + 0.003727065632119775, + 0.006154456175863743, + 0.0044197929091751575, + 0.010852945037186146, + 0.007551873568445444, + 0.015366208739578724, + 0.023466220125555992, + 0.006932857446372509, + 0.011061843484640121, + 0.01006957795470953, + 0.015367119573056698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.012539827264845371 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1892080456018448, + 0.12835252285003662, + 0.048936132341623306, + 0.178195983171463, + 0.12209736555814743, + 0.04064202308654785, + 0.08222906291484833, + 0.053658585995435715, + 0.05501611530780792, + 0.10166420787572861 + ], + "credit_map": [ + 0.01854378916323185, + 0.023474780842661858, + 0.027989240363240242, + 0.02859121933579445, + 0.021108541637659073, + 0.024184292182326317, + 0.022150026634335518, + 0.02316616103053093, + 0.011933032423257828, + 0.01578996330499649, + 0.012351103127002716, + 0.026825837790966034, + 0.011939141899347305, + 0.02003764733672142, + 0.014495130628347397, + 0.01498066633939743, + 0.005474923178553581, + 0.005572658497840166, + 0.006215947214514017, + 0.006789660081267357, + 0.005444211419671774, + 0.006510498467832804, + 0.005953766871243715, + 0.006974466145038605, + 0.017185168340802193, + 0.016951527446508408, + 0.023345720022916794, + 0.018939081579446793, + 0.022116228938102722, + 0.02683117613196373, + 0.03344569355249405, + 0.019381407648324966, + 0.010623279958963394, + 0.013544879853725433, + 0.01392277516424656, + 0.016374925151467323, + 0.014680384658277035, + 0.01524286437779665, + 0.023178592324256897, + 0.014529665932059288, + 0.005140369292348623, + 0.0053551229648292065, + 0.005997105967253447, + 0.005060581956058741, + 0.005298219621181488, + 0.004204436205327511, + 0.004390791058540344, + 0.005195395089685917, + 0.014867682009935379, + 0.006459843832999468, + 0.01428110059350729, + 0.011404522694647312, + 0.006216677837073803, + 0.006154896691441536, + 0.008448044769465923, + 0.01439629402011633, + 0.00676647387444973, + 0.007789261639118195, + 0.005463205743581057, + 0.011040477082133293, + 0.006917462684214115, + 0.0047224764712154865, + 0.005887198261916637, + 0.005072030704468489, + 0.007823070511221886, + 0.007705105468630791, + 0.008946347050368786, + 0.007327197585254908, + 0.006348893046379089, + 0.005204970017075539, + 0.0060630496591329575, + 0.005597485229372978, + 0.012545130215585232, + 0.0072551327757537365, + 0.012426387518644333, + 0.02550206147134304, + 0.008110174909234047, + 0.00892813690006733, + 0.010310444980859756, + 0.016586733981966972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26547521352767944, + 0.2854810357093811, + 0.04476853460073471, + 0.05790285766124725, + 0.04091246426105499, + 0.012957071885466576, + 0.11286398768424988, + 0.027965685352683067, + 0.02825905568897724, + 0.12341411411762238 + ], + "credit_map": [ + 0.02376622147858143, + 0.025453751906752586, + 0.03837505728006363, + 0.03360467404127121, + 0.034067731350660324, + 0.03632112964987755, + 0.03026881255209446, + 0.0436178594827652, + 0.019660858437418938, + 0.030252743512392044, + 0.027930626645684242, + 0.03132983669638634, + 0.026971952989697456, + 0.051357220858335495, + 0.06406649202108383, + 0.03391131013631821, + 0.004976509138941765, + 0.00397435063496232, + 0.005980700254440308, + 0.004981196019798517, + 0.004596758168190718, + 0.008711465634405613, + 0.004950667731463909, + 0.006596887484192848, + 0.003978752996772528, + 0.006881332956254482, + 0.008887121453881264, + 0.007879822514951229, + 0.0063942596316337585, + 0.008404534310102463, + 0.008267047815024853, + 0.007209985051304102, + 0.002478371374309063, + 0.004461022559553385, + 0.004985188599675894, + 0.005745600908994675, + 0.006052452605217695, + 0.005801254417747259, + 0.006423318292945623, + 0.004965256433933973, + 0.0011537219397723675, + 0.0013697859831154346, + 0.0013243111316114664, + 0.001845227088779211, + 0.0016706101596355438, + 0.0020459017250686884, + 0.0016069761477410793, + 0.0019405371276661754, + 0.01765337772667408, + 0.007154996506869793, + 0.024889754131436348, + 0.019438697025179863, + 0.0076728034764528275, + 0.008415181189775467, + 0.010918976739048958, + 0.016720207408070564, + 0.005299906246364117, + 0.0027069859206676483, + 0.004152465611696243, + 0.005320413038134575, + 0.003110573161393404, + 0.0019429126987233758, + 0.0029815062880516052, + 0.002450924599543214, + 0.0034509925171732903, + 0.0027880242560058832, + 0.0043470775708556175, + 0.005109081510454416, + 0.003603314748033881, + 0.002067364752292633, + 0.003540157340466976, + 0.003353043459355831, + 0.0197893176227808, + 0.012368536554276943, + 0.022525131702423096, + 0.01965128816664219, + 0.008659654296934605, + 0.012180319055914879, + 0.011412019841372967, + 0.016827844083309174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.005947629921138287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly tilted-up close-up captures the viewer's gaze firmly on the roar of a white tiger. Dramatic, high-contrast lighting cast from the side and above, sculpting every detail with pinpoint precision. The tiger's thick, snow-like fur is crisscrossed by inky black stripes that snake along its powerful facial features, strained with intense rage. Its mouth is stretched to its limit, a terrifying maw revealing a cavernous, dark pink interior. A massive, barbed tongue curls back in its roar, while ivory canines several inches long, dagger-sharp, ivory-colored canines gleam with a wet, cold gleam. A deafening roar erupts from deep within its throat, its wet black nose wrinkling violently in aggression, its thick white beard quivering with the force, and tiny drops of saliva, driven by the mighty sound, form a fleeting mist in the air. The background is completely blurred into an abstract block of dark green and brown, placing the undisputed focus on this face, full of primal power and endless ferocity, creating a terrifying and tense atmosphere where one can almost feel the air vibrate.", + "seed": 19644, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19644_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": -0.0061548032565042375, + "vlm_logprob": -0.0061548032565042375, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.0061548032565042375, + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0066131544299423695 + ], + "opposite_logprob_tokens": [ + -6.631613254547119 + ], + "logprob_reward": -0.0066131544299423695, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0066131544299423695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05671555921435356, + 0.16279122233390808, + 0.1986653357744217, + 0.06145566701889038, + 0.037960465997457504, + 0.08228456974029541, + 0.21353620290756226, + 0.10027555376291275, + 0.025935061275959015, + 0.06038035824894905 + ], + "credit_map": [ + 0.009245949797332287, + 0.0066503556445240974, + 0.009006974287331104, + 0.00882815569639206, + 0.005990855861455202, + 0.005903102457523346, + 0.006336199585348368, + 0.004753966350108385, + 0.04542672634124756, + 0.009917288087308407, + 0.05684477835893631, + 0.010725794360041618, + 0.008131306618452072, + 0.00608521793037653, + 0.010576663538813591, + 0.015083454549312592, + 0.019618650898337364, + 0.027573812752962112, + 0.01988893374800682, + 0.04133298620581627, + 0.024081643670797348, + 0.023906350135803223, + 0.024070244282484055, + 0.018192697316408157, + 0.006827356293797493, + 0.007128007244318724, + 0.005606635939329863, + 0.008916845545172691, + 0.009663729928433895, + 0.009736296720802784, + 0.0070923371240496635, + 0.006484461948275566, + 0.008083351887762547, + 0.002921514678746462, + 0.006859259679913521, + 0.0041314722038805485, + 0.0033594314008951187, + 0.0031118292827159166, + 0.004607520066201687, + 0.004886088892817497, + 0.01314224861562252, + 0.004495464731007814, + 0.020495476201176643, + 0.007787792012095451, + 0.007685428485274315, + 0.006051584612578154, + 0.013686138205230236, + 0.008940432220697403, + 0.039733387529850006, + 0.013940328732132912, + 0.013089701533317566, + 0.016145091503858566, + 0.016735611483454704, + 0.017588086426258087, + 0.07349876314401627, + 0.022805230692029, + 0.013036403805017471, + 0.011968939565122128, + 0.019053922966122627, + 0.024592045694589615, + 0.006726855877786875, + 0.005973811727017164, + 0.008935333229601383, + 0.009988244622945786, + 0.003740510903298855, + 0.0030617702286690474, + 0.001715100253932178, + 0.0024163895286619663, + 0.0037085802759975195, + 0.0030594158452004194, + 0.004739079624414444, + 0.003494214965030551, + 0.009010617621243, + 0.009153490886092186, + 0.004556776024401188, + 0.012800678610801697, + 0.007189759984612465, + 0.004624989349395037, + 0.007292627356946468, + 0.005751419346779585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.0066131544299423695 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019506252137944102 + ], + "opposite_logprob_tokens": [ + -6.251950740814209 + ], + "logprob_reward": -0.0019506252137944102, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019506252137944102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051942747086286545, + 0.08597055077552795, + 0.10690662264823914, + 0.1032932922244072, + 0.17786161601543427, + 0.2020244598388672, + 0.08914735913276672, + 0.057955339550971985, + 0.06126749888062477, + 0.06363046914339066 + ], + "credit_map": [ + 0.010566954500973225, + 0.005952552892267704, + 0.006256036460399628, + 0.005689265672117472, + 0.005167452618479729, + 0.005251749884337187, + 0.0066763609647750854, + 0.006382375955581665, + 0.019628198817372322, + 0.005874433554708958, + 0.016144275665283203, + 0.009141702204942703, + 0.006971090100705624, + 0.007656407076865435, + 0.012333032675087452, + 0.00822142232209444, + 0.011009618639945984, + 0.011224336922168732, + 0.013575766235589981, + 0.01889294572174549, + 0.01383224967867136, + 0.017525220289826393, + 0.012291043065488338, + 0.008555442094802856, + 0.01119452714920044, + 0.011682974174618721, + 0.010052438825368881, + 0.009487804025411606, + 0.018422506749629974, + 0.01409399975091219, + 0.015134477987885475, + 0.013224570080637932, + 0.03305250033736229, + 0.027471857145428658, + 0.030158812180161476, + 0.026634719222784042, + 0.013911253772675991, + 0.011921903118491173, + 0.015887804329395294, + 0.01882278360426426, + 0.035114213824272156, + 0.013761731795966625, + 0.04972562566399574, + 0.024055015295743942, + 0.016643909737467766, + 0.01579723320901394, + 0.026084305718541145, + 0.020842457190155983, + 0.013472647406160831, + 0.007271551992744207, + 0.006307477131485939, + 0.006902911234647036, + 0.008296545594930649, + 0.0067022377625107765, + 0.03157757222652435, + 0.008616425096988678, + 0.013686754740774632, + 0.006135650910437107, + 0.009172755293548107, + 0.005127947311848402, + 0.00481838546693325, + 0.0051016900688409805, + 0.006957086734473705, + 0.006955075077712536, + 0.00833094958215952, + 0.00604387978091836, + 0.007559496443718672, + 0.0077308546751737595, + 0.007457714062184095, + 0.008717313408851624, + 0.007636291440576315, + 0.007791007868945599, + 0.007680344395339489, + 0.007520332001149654, + 0.004286316689103842, + 0.012039738707244396, + 0.007711075246334076, + 0.007147486321628094, + 0.00916414987295866, + 0.00808103196322918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.0019506252137944102 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08905342221260071, + 0.11938221007585526, + 0.06693577021360397, + 0.037146586924791336, + 0.05469071865081787, + 0.09498291462659836, + 0.13112005591392517, + 0.09838887304067612, + 0.26100707054138184, + 0.04729241505265236 + ], + "credit_map": [ + 0.01938444748520851, + 0.009004448540508747, + 0.011755222454667091, + 0.009423933923244476, + 0.008344581350684166, + 0.010267031379044056, + 0.010379365645349026, + 0.010494393296539783, + 0.02961336448788643, + 0.008991939947009087, + 0.03385353088378906, + 0.00911911390721798, + 0.007772741839289665, + 0.006583636626601219, + 0.012749157845973969, + 0.010698715224862099, + 0.007555301766842604, + 0.007745049893856049, + 0.0060816192999482155, + 0.010693632066249847, + 0.006896750535815954, + 0.012289888225495815, + 0.008645877242088318, + 0.007027653511613607, + 0.00416135136038065, + 0.0038096383213996887, + 0.005218820180743933, + 0.004612825810909271, + 0.004320301581174135, + 0.005349484272301197, + 0.004306518007069826, + 0.005367650650441647, + 0.011038477532565594, + 0.005392958875745535, + 0.0074064177460968494, + 0.006084052380174398, + 0.005227821413427591, + 0.004784128163009882, + 0.005473430268466473, + 0.009283431805670261, + 0.01299531850963831, + 0.007964598946273327, + 0.014867396093904972, + 0.011859937570989132, + 0.00933749321848154, + 0.009840521961450577, + 0.016105016693472862, + 0.01201262604445219, + 0.022440992295742035, + 0.009656578302383423, + 0.010229725390672684, + 0.010834623128175735, + 0.011849206872284412, + 0.011712429113686085, + 0.03932127729058266, + 0.015075231902301311, + 0.01223254669457674, + 0.008456352166831493, + 0.021115314215421677, + 0.025250768288969994, + 0.006105002481490374, + 0.006877458188682795, + 0.010151546448469162, + 0.008199886418879032, + 0.03537945821881294, + 0.023298120126128197, + 0.030721597373485565, + 0.033612046390771866, + 0.0324675478041172, + 0.03705788403749466, + 0.03501316159963608, + 0.03345726802945137, + 0.004912207834422588, + 0.004764695651829243, + 0.002785366727039218, + 0.007080762647092342, + 0.005577025469392538, + 0.005776938982307911, + 0.009186179377138615, + 0.007209241855889559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.007625873666256666 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639340288937092 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639340288937092, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639340288937092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09842576086521149, + 0.12005547434091568, + 0.04892223700881004, + 0.07267478853464127, + 0.032419320195913315, + 0.12184146791696548, + 0.22806592285633087, + 0.05460687354207039, + 0.19387651979923248, + 0.029111599549651146 + ], + "credit_map": [ + 0.02457805722951889, + 0.010141043923795223, + 0.011961915530264378, + 0.009007547982037067, + 0.009080685675144196, + 0.011639246717095375, + 0.01219696644693613, + 0.009820306673645973, + 0.02903948910534382, + 0.008670836687088013, + 0.03056950680911541, + 0.012200826779007912, + 0.009437770582735538, + 0.008629098534584045, + 0.011259796097874641, + 0.01024814322590828, + 0.005009247921407223, + 0.009512302465736866, + 0.004354163538664579, + 0.00871657021343708, + 0.0048455726355314255, + 0.00648937514051795, + 0.005596369970589876, + 0.00439863558858633, + 0.009127062745392323, + 0.00580006605014205, + 0.0088313277810812, + 0.0074321250431239605, + 0.00888791773468256, + 0.009192959405481815, + 0.008827513083815575, + 0.014575819484889507, + 0.008199449628591537, + 0.0022335359826684, + 0.007490937598049641, + 0.003296493785455823, + 0.0025219060480594635, + 0.002138880779966712, + 0.0032938176300376654, + 0.0032442996744066477, + 0.028777368366718292, + 0.006770613137632608, + 0.04826606810092926, + 0.008516532368957996, + 0.0058668614365160465, + 0.008568312972784042, + 0.008186298422515392, + 0.0068894121795892715, + 0.03875941410660744, + 0.016129789873957634, + 0.012242908589541912, + 0.015098415315151215, + 0.016469880938529968, + 0.015576265752315521, + 0.09220051020383835, + 0.021588725969195366, + 0.010493073612451553, + 0.005260512698441744, + 0.005672273226082325, + 0.014136405661702156, + 0.003969812765717506, + 0.004542205482721329, + 0.005686532706022263, + 0.004846055526286364, + 0.02523394674062729, + 0.018105948343873024, + 0.023014117032289505, + 0.025882206857204437, + 0.023892642930150032, + 0.027914896607398987, + 0.024750499054789543, + 0.025082260370254517, + 0.0048698787577450275, + 0.003977373708039522, + 0.002483529970049858, + 0.003865230828523636, + 0.003693240461871028, + 0.0029250120278447866, + 0.00274082412943244, + 0.004556510131806135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.008639340288937092 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0662587434053421, + 0.35776662826538086, + 0.07020720839500427, + 0.07693249732255936, + 0.024743394926190376, + 0.057262565940618515, + 0.11934342235326767, + 0.037732597440481186, + 0.14422033727169037, + 0.045532625168561935 + ], + "credit_map": [ + 0.018707318231463432, + 0.00540006160736084, + 0.008460157550871372, + 0.006800659000873566, + 0.004928463138639927, + 0.005959746893495321, + 0.007385868579149246, + 0.008616466075181961, + 0.10370766371488571, + 0.02111676335334778, + 0.1305311769247055, + 0.01970278099179268, + 0.014771632850170135, + 0.013642833568155766, + 0.01789940893650055, + 0.036394376307725906, + 0.007867454551160336, + 0.011301782913506031, + 0.008508766070008278, + 0.010033873841166496, + 0.008042015135288239, + 0.011576053686439991, + 0.008235669694840908, + 0.004641597159206867, + 0.01057912316173315, + 0.006723887287080288, + 0.010390442796051502, + 0.010588083416223526, + 0.008231245912611485, + 0.010479011572897434, + 0.00692271301522851, + 0.013017992489039898, + 0.003022516379132867, + 0.002586815506219864, + 0.004919014405459166, + 0.0026413158047944307, + 0.0026878395583480597, + 0.002051521325483918, + 0.003435736522078514, + 0.0033986372873187065, + 0.013054125010967255, + 0.0033439178951084614, + 0.02203233540058136, + 0.004201047122478485, + 0.003009808948263526, + 0.0039622001349925995, + 0.004113492555916309, + 0.003545637708157301, + 0.018567688763141632, + 0.008133982308208942, + 0.005540729034692049, + 0.008704757317900658, + 0.011261167004704475, + 0.008751514367759228, + 0.04615144431591034, + 0.01223213691264391, + 0.008866438642144203, + 0.0030910891946405172, + 0.002612148877233267, + 0.00970136746764183, + 0.0031506726518273354, + 0.0036396284122020006, + 0.002838808810338378, + 0.0038324459455907345, + 0.020337745547294617, + 0.013684418052434921, + 0.019260412082076073, + 0.018990647047758102, + 0.01694166101515293, + 0.020485559478402138, + 0.016739975661039352, + 0.017779914662241936, + 0.005824803374707699, + 0.005756983533501625, + 0.005872000008821487, + 0.007777565158903599, + 0.005005695857107639, + 0.005020854528993368, + 0.005593584384769201, + 0.004681138787418604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly tilted-up close-up captures the viewer's gaze firmly on the roar of a white tiger. Dramatic, high-contrast lighting cast from the side and above, sculpting every detail with pinpoint precision. The tiger's thick, snow-like fur is crisscrossed by inky black stripes that snake along its powerful facial features, strained with intense rage. Its mouth is stretched to its limit, a terrifying maw revealing a cavernous, dark pink interior. A massive, barbed tongue curls back in its roar, while ivory canines several inches long, dagger-sharp, ivory-colored canines gleam with a wet, cold gleam. A deafening roar erupts from deep within its throat, its wet black nose wrinkling violently in aggression, its thick white beard quivering with the force, and tiny drops of saliva, driven by the mighty sound, form a fleeting mist in the air. The background is completely blurred into an abstract block of dark green and brown, placing the undisputed focus on this face, full of primal power and endless ferocity, creating a terrifying and tense atmosphere where one can almost feel the air vibrate.", + "seed": 19645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_2/index_3206_seed_19645_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": -0.006684685358777642, + "vlm_logprob": -0.006684685358777642, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.006684685358777642, + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005580797791481018 + ], + "opposite_logprob_tokens": [ + -7.005580902099609 + ], + "logprob_reward": -0.005580797791481018, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.005580797791481018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09410753101110458, + 0.03856859356164932, + 0.036897432059049606, + 0.16809938848018646, + 0.0590074397623539, + 0.34755250811576843, + 0.044979438185691833, + 0.06666059046983719, + 0.03997134789824486, + 0.10415571182966232 + ], + "credit_map": [ + 0.01454809121787548, + 0.008330930024385452, + 0.0119418203830719, + 0.007485877722501755, + 0.012345835566520691, + 0.006939306855201721, + 0.009223992936313152, + 0.023291686549782753, + 0.004739406984299421, + 0.002822715090587735, + 0.003991915378719568, + 0.007529865484684706, + 0.003917087335139513, + 0.0034467135556042194, + 0.00738291023299098, + 0.004737981129437685, + 0.005351558793336153, + 0.002260982058942318, + 0.00410411786288023, + 0.004066988360136747, + 0.004701895173639059, + 0.004341227933764458, + 0.003885337384417653, + 0.008185330778360367, + 0.02222234569489956, + 0.021026669070124626, + 0.0219928789883852, + 0.019969593733549118, + 0.017395244911313057, + 0.021797504276037216, + 0.02390693500638008, + 0.019788222387433052, + 0.0072187986224889755, + 0.005417732521891594, + 0.008197185583412647, + 0.004154866095632315, + 0.005298398435115814, + 0.0061457352712750435, + 0.007388172205537558, + 0.01518655102699995, + 0.039500243961811066, + 0.030747296288609505, + 0.03928200155496597, + 0.03383206948637962, + 0.029052702710032463, + 0.06828353554010391, + 0.0437188483774662, + 0.06313585489988327, + 0.005207403097301722, + 0.002734298352152109, + 0.008394655771553516, + 0.007954834029078484, + 0.006503199227154255, + 0.0033585475757718086, + 0.005878865253180265, + 0.00494763907045126, + 0.011004983447492123, + 0.004815962165594101, + 0.0106667997315526, + 0.009773215278983116, + 0.009091527201235294, + 0.004920657724142075, + 0.009560634382069111, + 0.006826814729720354, + 0.009152223356068134, + 0.0030509429052472115, + 0.005271154455840588, + 0.004935998469591141, + 0.004740442149341106, + 0.003036127658560872, + 0.004249230492860079, + 0.005535231437534094, + 0.018659932538866997, + 0.01044091209769249, + 0.011541122570633888, + 0.010432761162519455, + 0.01324742566794157, + 0.012149780057370663, + 0.013439565896987915, + 0.014244217425584793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.005580797791481018 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028297165408730507 + ], + "opposite_logprob_tokens": [ + -5.877829551696777 + ], + "logprob_reward": -0.0028297165408730507, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028297165408730507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23313678801059723, + 0.14563453197479248, + 0.03687542304396629, + 0.07090993970632553, + 0.09236110001802444, + 0.1348056048154831, + 0.058172184973955154, + 0.03751973807811737, + 0.024423694238066673, + 0.16616098582744598 + ], + "credit_map": [ + 0.03878946602344513, + 0.023948613554239273, + 0.05525173246860504, + 0.018676863983273506, + 0.01994466595351696, + 0.025586608797311783, + 0.027425644919276237, + 0.023513199761509895, + 0.016714543104171753, + 0.01222852524369955, + 0.01523477490991354, + 0.04601163789629936, + 0.01907995715737343, + 0.011709741316735744, + 0.01394644659012556, + 0.01070892158895731, + 0.0063550123013556, + 0.0028107918333262205, + 0.0046968236565589905, + 0.0038453538436442614, + 0.004018433392047882, + 0.003747235983610153, + 0.0045970953069627285, + 0.0068046776577830315, + 0.008102369494736195, + 0.009699930436909199, + 0.009424987249076366, + 0.0060737417079508305, + 0.0071249003522098064, + 0.009682552888989449, + 0.010917405597865582, + 0.009884052909910679, + 0.013682469725608826, + 0.008360285311937332, + 0.010321928188204765, + 0.007468125317245722, + 0.010008520446717739, + 0.009389708749949932, + 0.00955791212618351, + 0.023572156205773354, + 0.015937095507979393, + 0.01199311949312687, + 0.015111232176423073, + 0.0143108069896698, + 0.02427051216363907, + 0.012697494588792324, + 0.022611333057284355, + 0.017874017357826233, + 0.009566009044647217, + 0.0034592195879667997, + 0.01423192210495472, + 0.006252821534872055, + 0.008416477590799332, + 0.0038145517464727163, + 0.00620680395513773, + 0.0062243822030723095, + 0.005663735792040825, + 0.0031855099368840456, + 0.005169873125851154, + 0.005059842020273209, + 0.0048914444632828236, + 0.0031052413396537304, + 0.00449790246784687, + 0.0059461938217282295, + 0.006049443501979113, + 0.0016344337491318583, + 0.0031038406305015087, + 0.0025925193913280964, + 0.0034118634648621082, + 0.002071568975225091, + 0.003276700619608164, + 0.0022833270486444235, + 0.01813138835132122, + 0.015956608578562737, + 0.02081902138888836, + 0.015130993910133839, + 0.037746138870716095, + 0.01889721117913723, + 0.018558649346232414, + 0.02092098630964756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0028297165408730507 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0822029709815979, + 0.05775723233819008, + 0.035691406577825546, + 0.09088566899299622, + 0.055448222905397415, + 0.29204773902893066, + 0.09623068571090698, + 0.09003406763076782, + 0.04116465151309967, + 0.1585373729467392 + ], + "credit_map": [ + 0.010943460278213024, + 0.007295635063201189, + 0.016344061121344566, + 0.008172704838216305, + 0.014157320372760296, + 0.005952346138656139, + 0.011898275464773178, + 0.007439168635755777, + 0.007883955724537373, + 0.004136885050684214, + 0.009393449872732162, + 0.007164811249822378, + 0.007182398345321417, + 0.005141095723956823, + 0.009215379133820534, + 0.00763925164937973, + 0.0038282726891338825, + 0.0030231145210564137, + 0.004123538266867399, + 0.004810929764062166, + 0.0038974154740571976, + 0.004532407037913799, + 0.0035812139976769686, + 0.007894515991210938, + 0.00952872447669506, + 0.008484707213938236, + 0.010642759501934052, + 0.011392763815820217, + 0.012332391925156116, + 0.007597885560244322, + 0.015539144165813923, + 0.015367291867733002, + 0.009977875277400017, + 0.005221866071224213, + 0.007180104963481426, + 0.0037516646552830935, + 0.004298936575651169, + 0.004878204315900803, + 0.005759911146014929, + 0.014379660598933697, + 0.034569427371025085, + 0.025105463340878487, + 0.03857957944273949, + 0.02778610959649086, + 0.030677758157253265, + 0.05632387474179268, + 0.03517479822039604, + 0.04383072257041931, + 0.011965676210820675, + 0.006868110504001379, + 0.02156258560717106, + 0.009041793644428253, + 0.018209949135780334, + 0.006677867379039526, + 0.01227071788161993, + 0.0096339862793684, + 0.01293848268687725, + 0.007998939603567123, + 0.015255597420036793, + 0.010893982835114002, + 0.014917894266545773, + 0.007055558729916811, + 0.010631181299686432, + 0.010342428460717201, + 0.00731396209448576, + 0.0025370460934937, + 0.004967689048498869, + 0.004762120079249144, + 0.005853690672665834, + 0.004006071016192436, + 0.008553341962397099, + 0.0031707307789474726, + 0.03341410309076309, + 0.013418575748801231, + 0.019947875291109085, + 0.015162594616413116, + 0.016038091853260994, + 0.017292220145463943, + 0.018197450786828995, + 0.02506645955145359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.007626701612025499 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06942905485630035, + 0.05828084796667099, + 0.050335634499788284, + 0.4105721414089203, + 0.031437452882528305, + 0.21458999812602997, + 0.031200211495161057, + 0.05512262508273125, + 0.029768643900752068, + 0.0492633618414402 + ], + "credit_map": [ + 0.00963644403964281, + 0.005578306503593922, + 0.015489649027585983, + 0.007111800368875265, + 0.011138652451336384, + 0.005201537162065506, + 0.008424798958003521, + 0.006847861688584089, + 0.0070923506282269955, + 0.005105968564748764, + 0.01280496921390295, + 0.008444040082395077, + 0.004472974222153425, + 0.005522218532860279, + 0.008468321524560452, + 0.0063700079917907715, + 0.01104834582656622, + 0.0037247459404170513, + 0.007486029528081417, + 0.005616422276943922, + 0.004813202656805515, + 0.004925246816128492, + 0.004024265334010124, + 0.008697379380464554, + 0.05002830550074577, + 0.052011728286743164, + 0.056014284491539, + 0.046623341739177704, + 0.04423588514328003, + 0.05509510263800621, + 0.057826098054647446, + 0.04873739928007126, + 0.007134214974939823, + 0.00421525863930583, + 0.003590419190004468, + 0.0027886470779776573, + 0.003199509112164378, + 0.0024926315527409315, + 0.003587278537452221, + 0.004429493099451065, + 0.022608788684010506, + 0.01742498390376568, + 0.025265496224164963, + 0.01987471804022789, + 0.018730856478214264, + 0.04601054638624191, + 0.026308031752705574, + 0.038366567343473434, + 0.004414067137986422, + 0.0026134788058698177, + 0.006768631283193827, + 0.003360940609127283, + 0.005906655918806791, + 0.0020320608746260405, + 0.003520007710903883, + 0.0025843693874776363, + 0.011573480442166328, + 0.0036782228853553534, + 0.00674678198993206, + 0.0071489582769572735, + 0.006437448784708977, + 0.00469072675332427, + 0.006631499156355858, + 0.008215502835810184, + 0.00519346771761775, + 0.0021509325597435236, + 0.004190271254628897, + 0.00435492442920804, + 0.0035875332541763783, + 0.0021227861288934946, + 0.004200269468128681, + 0.0039684586226940155, + 0.00722799077630043, + 0.005875522270798683, + 0.004916556645184755, + 0.003766051260754466, + 0.009267274290323257, + 0.006483821198344231, + 0.004799848422408104, + 0.00692629674449563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.014194712042808533 + }, + { + "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 static, slightly tilted-up close-up of a white tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's fur white and thick, with inky black stripes on its face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's mouth open wide, revealing a dark pink interior, long ivory canines, and a massive barbed tongue curling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting dramatic and high-contrast, coming from the side and above to sculpt the tiger's features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into an abstract block of dark green and brown?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609877049922943, + 0.05473644286394119, + 0.023018477484583855, + 0.19196173548698425, + 0.053655240684747696, + 0.3024824559688568, + 0.04247033968567848, + 0.05345745384693146, + 0.02380300499498844, + 0.0934271514415741 + ], + "credit_map": [ + 0.023540759459137917, + 0.017699331045150757, + 0.03123643435537815, + 0.014117961749434471, + 0.01347680576145649, + 0.020701251924037933, + 0.016014764085412025, + 0.024200405925512314, + 0.008587566204369068, + 0.004047479014843702, + 0.0073153553530573845, + 0.015767257660627365, + 0.004536001943051815, + 0.005013688001781702, + 0.0035932462196797132, + 0.0058758556842803955, + 0.005857572425156832, + 0.0019355141557753086, + 0.0032466244883835316, + 0.0022140045184642076, + 0.0022439463064074516, + 0.0018055454129353166, + 0.002508686389774084, + 0.003206586465239525, + 0.02280542440712452, + 0.02213590033352375, + 0.026509683579206467, + 0.02203897200524807, + 0.02261294610798359, + 0.02513176202774048, + 0.028798244893550873, + 0.0219288133084774, + 0.011126513592898846, + 0.004906777758151293, + 0.006598659325391054, + 0.003179729450494051, + 0.0037163032684475183, + 0.0051480005495250225, + 0.006342217326164246, + 0.012637040577828884, + 0.037190746515989304, + 0.024158407002687454, + 0.036652401089668274, + 0.02846592478454113, + 0.025583334267139435, + 0.05743755027651787, + 0.03839056193828583, + 0.05460352450609207, + 0.00813413504511118, + 0.004339156672358513, + 0.006687853019684553, + 0.004909084644168615, + 0.005735136102885008, + 0.003307987004518509, + 0.004754632245749235, + 0.004602361470460892, + 0.015798253938555717, + 0.0038948345463722944, + 0.007882063277065754, + 0.006251848768442869, + 0.005908007733523846, + 0.0036769642028957605, + 0.004230036865919828, + 0.005815449170768261, + 0.005162564571946859, + 0.001727065071463585, + 0.00295400177128613, + 0.0026609518099576235, + 0.0032663217280060053, + 0.0018580126343294978, + 0.00336753879673779, + 0.00280655175447464, + 0.019077440723776817, + 0.011161856353282928, + 0.011713717132806778, + 0.009834570810198784, + 0.009132900275290012, + 0.00845915637910366, + 0.008984791114926338, + 0.015062724240124226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.0031914988067001104 + } + ], + "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/0000167/group_3/index_tensor([7219, 7219])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_tensor([7219, 7219])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d0f47cd43cf5ab748cb2a14dab6d2597727bfa64 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_tensor([7219, 7219])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sharp, dynamic swooping shot zooms in on a futuristic electronic door lock mounted on a heavy metal door. Just as the camera is about to make contact, the entire frame switches to an extreme high-speed photography style, as if time has frozen. Crafted from smooth, brushed black titanium, the lock, devoid of physical buttons and marked only by a slowly breathing, ring of eerie blue LED lights, suddenly explodes from within, triggered by an invisible force. The lock violently explodes from the center outward, sending countless gleaming metal fragments, delicate internal gears, glowing microcircuits, and delicate copper wires radiating through the air in a magnificent dance of destruction. Each swirling fragment distinctly reflects the cold white light of the surrounding environment, while tiny arcs of blue electricity flicker between the fractured components, stretching the silent explosion into a visual spectacle of technological beauty and violent impact, creating a breathtaking atmosphere of tension where the moment feels like eternity.", + "seed": 24695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24695_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": -0.026737093221163376, + "vlm_logprob": -0.026737093221163376, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.026737093221163376, + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04972269386053085 + ], + "opposite_logprob_tokens": [ + -3.174722671508789 + ], + "logprob_reward": -0.04972269386053085, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04972269386053085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2389751821756363, + 0.346626341342926, + 0.041507307440042496, + 0.04652628302574158, + 0.048170752823352814, + 0.018535243347287178, + 0.12351223081350327, + 0.043513353914022446, + 0.03509864956140518, + 0.057534631341695786 + ], + "credit_map": [ + 0.016487425193190575, + 0.007424397859722376, + 0.027080630883574486, + 0.011775610968470573, + 0.007586638908833265, + 0.08287354558706284, + 0.011465338058769703, + 0.07428160309791565, + 0.018871068954467773, + 0.009473382495343685, + 0.11250163614749908, + 0.018147435039281845, + 0.008419672958552837, + 0.13639506697654724, + 0.013397784903645515, + 0.02942027524113655, + 0.00524706207215786, + 0.004100405611097813, + 0.005224402993917465, + 0.0066811703145504, + 0.0039617582224309444, + 0.004549480974674225, + 0.006710267625749111, + 0.005032759625464678, + 0.005349248647689819, + 0.003938704263418913, + 0.004653812386095524, + 0.005115675274282694, + 0.004048175644129515, + 0.007143849972635508, + 0.005919190123677254, + 0.010357627645134926, + 0.005895473528653383, + 0.0051544164307415485, + 0.004442771431058645, + 0.004432193003594875, + 0.006927470210939646, + 0.007299910299479961, + 0.007972117513418198, + 0.006046405527740717, + 0.0023761396296322346, + 0.0016638905508443713, + 0.001973393140360713, + 0.002544837072491646, + 0.0020074413623660803, + 0.002457764931023121, + 0.0029957627411931753, + 0.0025160156656056643, + 0.04281371831893921, + 0.004706528037786484, + 0.0076874191872775555, + 0.0056510670110583305, + 0.004054222721606493, + 0.00667292857542634, + 0.011490887962281704, + 0.040435463190078735, + 0.005852566100656986, + 0.004880599211901426, + 0.0055670905858278275, + 0.00538954883813858, + 0.0039212810806930065, + 0.007207484450191259, + 0.00497285695746541, + 0.005721932277083397, + 0.002595691941678524, + 0.0032153946813195944, + 0.0035470579750835896, + 0.005065628327429295, + 0.006842398084700108, + 0.004161246120929718, + 0.0037789775524288416, + 0.005892259068787098, + 0.004838231950998306, + 0.00380198797211051, + 0.013112811371684074, + 0.007386975921690464, + 0.005889169406145811, + 0.010307441465556622, + 0.004686415195465088, + 0.007511602248996496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04972269386053085 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898569852113724 + ], + "opposite_logprob_tokens": [ + -2.5789856910705566 + ], + "logprob_reward": -0.07898569852113724, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898569852113724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08883205056190491, + 0.7009645104408264, + 0.02168915420770645, + 0.02472706139087677, + 0.022904422134160995, + 0.016938360407948494, + 0.049026843160390854, + 0.014870264567434788, + 0.012180394493043423, + 0.047866933047771454 + ], + "credit_map": [ + 0.0073574078269302845, + 0.0025453136768192053, + 0.01429956778883934, + 0.004237625282257795, + 0.002612758195027709, + 0.02389535866677761, + 0.006001617293804884, + 0.027882404625415802, + 0.03506904095411301, + 0.016289135441184044, + 0.24447710812091827, + 0.03383645787835121, + 0.015464067459106445, + 0.2822355031967163, + 0.022165263071656227, + 0.051427919417619705, + 0.002925748471170664, + 0.0023317779414355755, + 0.0023973046336323023, + 0.0028040155302733183, + 0.0023147687315940857, + 0.0033348200377076864, + 0.002751880092546344, + 0.0028288387693464756, + 0.003544151084497571, + 0.0017441860400140285, + 0.002977559342980385, + 0.001941813388839364, + 0.0018905134638771415, + 0.00443005608394742, + 0.002191732870414853, + 0.006007049698382616, + 0.0022204227279871702, + 0.001924629439599812, + 0.0016390839591622353, + 0.002632502233609557, + 0.003305352758616209, + 0.0035776656586676836, + 0.004693720489740372, + 0.0029110440518707037, + 0.0018733632750809193, + 0.0017244841437786818, + 0.0017715564463287592, + 0.0018947962671518326, + 0.001466649933718145, + 0.001907296827994287, + 0.0036316646728664637, + 0.0026685481425374746, + 0.01269543543457985, + 0.0020118325483053923, + 0.00344702391885221, + 0.0032000686042010784, + 0.0027942201122641563, + 0.003949132282286882, + 0.007663509342819452, + 0.013265622779726982, + 0.0015575917204841971, + 0.0019058112520724535, + 0.0019333529053255916, + 0.0019054845906794071, + 0.0015073016984388232, + 0.0018764291889965534, + 0.002133400412276387, + 0.0020508929155766964, + 0.0009507578215561807, + 0.0010482070501893759, + 0.0015972699038684368, + 0.0015752444742247462, + 0.0018777084769681096, + 0.0016598900547251105, + 0.0014735935255885124, + 0.001997723476961255, + 0.003912602085620165, + 0.0029978863894939423, + 0.01081724464893341, + 0.005440519191324711, + 0.006163980346173048, + 0.008831357583403587, + 0.00357497064396739, + 0.006128372624516487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.07898569852113724 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009235645993612707 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009235645993612707, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009235645993612707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13057903945446014, + 0.625616729259491, + 0.02925332449376583, + 0.038531508296728134, + 0.027597181499004364, + 0.019996201619505882, + 0.025184348225593567, + 0.009536209516227245, + 0.08148204535245895, + 0.01222345419228077 + ], + "credit_map": [ + 0.011901618912816048, + 0.008702828548848629, + 0.027955079451203346, + 0.009291641414165497, + 0.008360021747648716, + 0.036876287311315536, + 0.010243680328130722, + 0.017247872427105904, + 0.02831261418759823, + 0.01600973680615425, + 0.1385575234889984, + 0.02647174708545208, + 0.014887060970067978, + 0.257392942905426, + 0.02167062647640705, + 0.12231440097093582, + 0.003915600012987852, + 0.002435460453853011, + 0.005045012105256319, + 0.004382470156997442, + 0.002135717775672674, + 0.0035596031229943037, + 0.004068789072334766, + 0.0037106683012098074, + 0.006141018122434616, + 0.0027430946938693523, + 0.0042853401973843575, + 0.0034936154261231422, + 0.0030502816662192345, + 0.006669895723462105, + 0.003437501611188054, + 0.008710757829248905, + 0.0031774851959198713, + 0.002677129814401269, + 0.002428290667012334, + 0.0030055735260248184, + 0.0038795853033661842, + 0.003936655819416046, + 0.005388158839195967, + 0.003104301169514656, + 0.002314398530870676, + 0.0024625901132822037, + 0.0027561874594539404, + 0.0026601748540997505, + 0.0017428124556317925, + 0.0026148895267397165, + 0.00272921659052372, + 0.0027159310411661863, + 0.006097420118749142, + 0.0016518428456038237, + 0.002503756433725357, + 0.0015658396296203136, + 0.0017981628188863397, + 0.002016057027503848, + 0.0038581835106015205, + 0.005693082232028246, + 0.0008691422990523279, + 0.0010038058971986175, + 0.001160990446805954, + 0.0011754399165511131, + 0.0009658196358941495, + 0.0011803379748016596, + 0.0014726229710504413, + 0.0017080488614737988, + 0.005286733154207468, + 0.0028273649513721466, + 0.024853304028511047, + 0.006016107276082039, + 0.007561737205833197, + 0.017936566844582558, + 0.006496561225503683, + 0.01050365436822176, + 0.0013166177086532116, + 0.0010925035458058119, + 0.001257590251043439, + 0.0013743877643719316, + 0.0027620422188192606, + 0.0015776952495798469, + 0.0015788184246048331, + 0.0012637977488338947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0009235645993612707 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004378790326882154 + ], + "opposite_logprob_tokens": [ + -7.7504377365112305 + ], + "logprob_reward": -0.0004378790326882154, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004378790326882154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24368901550769806, + 0.43643811345100403, + 0.028194395825266838, + 0.04055757075548172, + 0.03277977928519249, + 0.03685115650296211, + 0.047854404896497726, + 0.05444677919149399, + 0.03962258622050285, + 0.03956618160009384 + ], + "credit_map": [ + 0.02184591256082058, + 0.008698497898876667, + 0.03804656118154526, + 0.01289624348282814, + 0.0065275877714157104, + 0.06435109674930573, + 0.015182573348283768, + 0.07614053785800934, + 0.021278085187077522, + 0.009970881044864655, + 0.1672181785106659, + 0.021227993071079254, + 0.009293857030570507, + 0.16164004802703857, + 0.012734395451843739, + 0.03307469189167023, + 0.002553851343691349, + 0.004140292759984732, + 0.0031560214702039957, + 0.0035864978563040495, + 0.004205915145576, + 0.003989088349044323, + 0.0035969207528978586, + 0.0029658067505806684, + 0.004261508118361235, + 0.004018602892756462, + 0.004067079164087772, + 0.004628216382116079, + 0.0061195241287350655, + 0.005081950221210718, + 0.007009935565292835, + 0.005370755214244127, + 0.0037958768662065268, + 0.003300196025520563, + 0.0031257716473191977, + 0.003824846586212516, + 0.0038874957244843245, + 0.005259128287434578, + 0.005240131635218859, + 0.004346332512795925, + 0.00376156996935606, + 0.003109340788796544, + 0.0050202785059809685, + 0.004860835149884224, + 0.003815797856077552, + 0.006361616775393486, + 0.005470946431159973, + 0.004450772423297167, + 0.004819896537810564, + 0.0036801956593990326, + 0.004323558881878853, + 0.0049003977328538895, + 0.005358283407986164, + 0.004752371925860643, + 0.012116638012230396, + 0.00790305808186531, + 0.006195043213665485, + 0.006698952987790108, + 0.007012458052486181, + 0.006471136584877968, + 0.005969919729977846, + 0.007757560815662146, + 0.00677452189847827, + 0.0075671919621527195, + 0.00356367789208889, + 0.0040148100815713406, + 0.005695333704352379, + 0.0043898154981434345, + 0.005611912813037634, + 0.007134860381484032, + 0.005120505578815937, + 0.004091671202331781, + 0.0032751644030213356, + 0.003083711490035057, + 0.007699351757764816, + 0.004732654895633459, + 0.005783458240330219, + 0.006973454728722572, + 0.0034656873904168606, + 0.0045526945032179356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0004378790326882154 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615630092099309 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.003615630092099309, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615630092099309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12914051115512848, + 0.6035915017127991, + 0.02105647139251232, + 0.01996052823960781, + 0.01471851859241724, + 0.01840609684586525, + 0.13821513950824738, + 0.02464858628809452, + 0.014734677970409393, + 0.015527976676821709 + ], + "credit_map": [ + 0.009675632230937481, + 0.0035399473272264004, + 0.014341835863888264, + 0.005633471067994833, + 0.0037075744476169348, + 0.04207416996359825, + 0.007061546202749014, + 0.043106336146593094, + 0.02969745174050331, + 0.014078023843467236, + 0.20712615549564362, + 0.02869170904159546, + 0.013112029060721397, + 0.24477142095565796, + 0.01925606280565262, + 0.04685862362384796, + 0.0019432340050116181, + 0.0023227070923894644, + 0.00200619432143867, + 0.0035170295741409063, + 0.002648821799084544, + 0.0027302131056785583, + 0.003364028874784708, + 0.00252424250356853, + 0.0023062003310769796, + 0.0019299748819321394, + 0.002039714017882943, + 0.0021129888482391834, + 0.0024057128466665745, + 0.003071559127420187, + 0.0027284927200526, + 0.0033658845350146294, + 0.001551185268908739, + 0.0013204643037170172, + 0.0012895951513200998, + 0.00183633784763515, + 0.0016830272506922483, + 0.0024528412614017725, + 0.0026662019081413746, + 0.0019188658334314823, + 0.002322568790987134, + 0.001968441531062126, + 0.002309990581125021, + 0.0020782225765287876, + 0.001663166331127286, + 0.0024272368755191565, + 0.002866320312023163, + 0.0027701493818312883, + 0.050129637122154236, + 0.00517340749502182, + 0.008631812408566475, + 0.004092007409781218, + 0.0033822120167315006, + 0.0081187067553401, + 0.011107886210083961, + 0.04757946729660034, + 0.0023090566974133253, + 0.0021139017771929502, + 0.0035907949786633253, + 0.0032152559142559767, + 0.00237684091553092, + 0.004179623909294605, + 0.003528150962665677, + 0.0033349625300616026, + 0.0014052990591153502, + 0.0010971430456265807, + 0.002222517505288124, + 0.0017110779881477356, + 0.0024908562190830708, + 0.002098418539389968, + 0.0015274620382115245, + 0.0021819027606397867, + 0.0013810349628329277, + 0.0014058856759220362, + 0.002756810514256358, + 0.002170637948438525, + 0.002264623297378421, + 0.0021321375388652086, + 0.0018552846740931273, + 0.001561561948619783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.003615630092099309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot zooms in on a futuristic electronic door lock mounted on a heavy metal door. Just as the camera is about to make contact, the entire frame switches to an extreme high-speed photography style, as if time has frozen. Crafted from smooth, brushed black titanium, the lock, devoid of physical buttons and marked only by a slowly breathing, ring of eerie blue LED lights, suddenly explodes from within, triggered by an invisible force. The lock violently explodes from the center outward, sending countless gleaming metal fragments, delicate internal gears, glowing microcircuits, and delicate copper wires radiating through the air in a magnificent dance of destruction. Each swirling fragment distinctly reflects the cold white light of the surrounding environment, while tiny arcs of blue electricity flicker between the fractured components, stretching the silent explosion into a visual spectacle of technological beauty and violent impact, creating a breathtaking atmosphere of tension where the moment feels like eternity.", + "seed": 24696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24696_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": -0.0462701638112776, + "vlm_logprob": -0.0462701638112776, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.0462701638112776, + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0188209880143404 + ], + "opposite_logprob_tokens": [ + -4.268820762634277 + ], + "logprob_reward": -0.0188209880143404, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0188209880143404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040877655148506165, + 0.01596471108496189, + 0.08530820161104202, + 0.5990349054336548, + 0.014440752565860748, + 0.01397268008440733, + 0.011542200110852718, + 0.11835185438394547, + 0.08777237683534622, + 0.012734596617519855 + ], + "credit_map": [ + 0.00425384659320116, + 0.0038904997054487467, + 0.004148060921579599, + 0.006357069127261639, + 0.0029908569995313883, + 0.007202750537544489, + 0.005664717871695757, + 0.006369858514517546, + 0.0016808697255328298, + 0.00091921555576846, + 0.0011289120884612203, + 0.0016720074927434325, + 0.0008408586145378649, + 0.004496697336435318, + 0.0017719556344673038, + 0.003454196033999324, + 0.0022069138940423727, + 0.0009812697535380721, + 0.0013353971298784018, + 0.0050824047066271305, + 0.002527831122279167, + 0.03317898511886597, + 0.0050107636488974094, + 0.03498465195298195, + 0.08530530333518982, + 0.02711627259850502, + 0.048965953290462494, + 0.053490109741687775, + 0.03755415603518486, + 0.10805386304855347, + 0.06571308523416519, + 0.17283622920513153, + 0.001741801854223013, + 0.0014967942843213677, + 0.002295442158356309, + 0.0014813154703006148, + 0.0010803879704326391, + 0.00201448705047369, + 0.0018470492213964462, + 0.0024834766518324614, + 0.002037568250671029, + 0.0009678594651632011, + 0.0024951596278697252, + 0.001231153029948473, + 0.001119008637033403, + 0.0020074567291885614, + 0.001437544939108193, + 0.0026769302785396576, + 0.0018343006959185004, + 0.0013732414226979017, + 0.0017544878646731377, + 0.0016432347474619746, + 0.0011308916145935655, + 0.0013420706382021308, + 0.0011564517626538873, + 0.001307522295974195, + 0.01500528771430254, + 0.01467890478670597, + 0.015485401265323162, + 0.008809315040707588, + 0.010951771400868893, + 0.01596192456781864, + 0.009961689822375774, + 0.027497565373778343, + 0.00772476801648736, + 0.013002484105527401, + 0.006929165683686733, + 0.008831250481307507, + 0.022308440878987312, + 0.013438452035188675, + 0.009785005822777748, + 0.005752818658947945, + 0.0009380136034451425, + 0.0013379775919020176, + 0.0013683967990800738, + 0.0016065033851191401, + 0.002512051723897457, + 0.0017534877406433225, + 0.0017271146643906832, + 0.0014910524478182197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.0188209880143404 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154041051864624 + ], + "opposite_logprob_tokens": [ + -1.701540470123291 + ], + "logprob_reward": -0.20154041051864624, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154041051864624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058309029787778854, + 0.030567225068807602, + 0.09413401037454605, + 0.6965481042861938, + 0.013975273817777634, + 0.008483846671879292, + 0.017849361523985863, + 0.008390182629227638, + 0.02843778021633625, + 0.04330514371395111 + ], + "credit_map": [ + 0.0060984971933066845, + 0.004545985721051693, + 0.0044777002185583115, + 0.01037459634244442, + 0.006183486431837082, + 0.008956056088209152, + 0.007971060462296009, + 0.009701649658381939, + 0.0026131027843803167, + 0.0015243124216794968, + 0.0019233418861404061, + 0.0026121255941689014, + 0.002139126183465123, + 0.007937921211123466, + 0.0030996566638350487, + 0.008717641234397888, + 0.003099547466263175, + 0.002087136497721076, + 0.00255905045196414, + 0.00691742729395628, + 0.003815923584625125, + 0.03407179191708565, + 0.006141824647784233, + 0.035441309213638306, + 0.09347651153802872, + 0.03014376014471054, + 0.056155722588300705, + 0.061470549553632736, + 0.04412204027175903, + 0.12929221987724304, + 0.07831919938325882, + 0.20356814563274384, + 0.0013578423531726003, + 0.0010116422781720757, + 0.0021595864091068506, + 0.001288514118641615, + 0.0010711128124967217, + 0.002777363173663616, + 0.0012662258232012391, + 0.003042986849322915, + 0.0020060467068105936, + 0.0005905759753659368, + 0.0011284577194601297, + 0.0007335201371461153, + 0.0007023989455774426, + 0.0011785476235672832, + 0.00080001016613096, + 0.001344290329143405, + 0.004123882856220007, + 0.0018680626526474953, + 0.003557636635378003, + 0.0013290229253470898, + 0.0017888231668621302, + 0.0018019827548414469, + 0.0013510201824828982, + 0.0020289309322834015, + 0.0012069221120327711, + 0.001112783094868064, + 0.0010338154388591647, + 0.0009507344220764935, + 0.000761670817155391, + 0.0012754816561937332, + 0.0007061513024382293, + 0.001342623378150165, + 0.002311199437826872, + 0.0040502846240997314, + 0.0022501088678836823, + 0.0031659845262765884, + 0.007446120958775282, + 0.0044418261386454105, + 0.00322609581053257, + 0.0015461576404049993, + 0.003357044653967023, + 0.004437299910932779, + 0.007287151645869017, + 0.007240660488605499, + 0.006007654592394829, + 0.006229241844266653, + 0.004743112251162529, + 0.0040029799565672874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.20154041051864624 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06085722893476486, + 0.020219968631863594, + 0.0994611456990242, + 0.6647118926048279, + 0.03223949670791626, + 0.011625880375504494, + 0.011821683496236801, + 0.048866450786590576, + 0.03296543285250664, + 0.017230818048119545 + ], + "credit_map": [ + 0.0072293709963560104, + 0.0050956010818481445, + 0.004739988595247269, + 0.010350584983825684, + 0.0060497308149933815, + 0.01105562411248684, + 0.007351768668740988, + 0.008984560146927834, + 0.0017271337565034628, + 0.0014752275310456753, + 0.0012623707298189402, + 0.0026544893626123667, + 0.0021120544988662004, + 0.00450777867808938, + 0.0022999814245849848, + 0.004180933348834515, + 0.0026114792563021183, + 0.0011583301238715649, + 0.0016889183316379786, + 0.005687878001481295, + 0.002862250432372093, + 0.03895137086510658, + 0.0058533730916678905, + 0.040647540241479874, + 0.08907691389322281, + 0.02920467220246792, + 0.05349764972925186, + 0.058068159967660904, + 0.04177594929933548, + 0.12372156977653503, + 0.07474986463785172, + 0.19461709260940552, + 0.003884568577632308, + 0.0024342720862478018, + 0.006850411184132099, + 0.0027419286780059338, + 0.003061567433178425, + 0.0051713502034544945, + 0.003753181779757142, + 0.004342219792306423, + 0.0026097670197486877, + 0.0008814952452667058, + 0.0018616425804793835, + 0.0008993259398266673, + 0.0008084587752819061, + 0.0016010133549571037, + 0.00095599569613114, + 0.0020081817638128996, + 0.002493416192010045, + 0.0014130487106740475, + 0.0018073554383590817, + 0.0007520780200138688, + 0.001637288718484342, + 0.0018474472453817725, + 0.0007541376398876309, + 0.001116911880671978, + 0.006481481716036797, + 0.006076094228774309, + 0.006259521935135126, + 0.00419781357049942, + 0.00456305081024766, + 0.005770835094153881, + 0.004211847670376301, + 0.011305807158350945, + 0.002609776332974434, + 0.00500286091119051, + 0.0026401143986731768, + 0.003495223820209503, + 0.00858985260128975, + 0.005058096721768379, + 0.0037467586807906628, + 0.001822749967686832, + 0.001219628844410181, + 0.0020509378518909216, + 0.001843765377998352, + 0.002302506472915411, + 0.0036948686465620995, + 0.00261670327745378, + 0.0016758573474362493, + 0.0018265494145452976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.008640285581350327 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006344689172692597 + ], + "opposite_logprob_tokens": [ + -7.375634670257568 + ], + "logprob_reward": -0.0006344689172692597, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006344689172692597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07177231460809708, + 0.039759986102581024, + 0.06019502505660057, + 0.57393878698349, + 0.025303885340690613, + 0.06075766682624817, + 0.018027188256382942, + 0.03827248513698578, + 0.07080760598182678, + 0.04116502404212952 + ], + "credit_map": [ + 0.007577887270599604, + 0.006567094009369612, + 0.005761316977441311, + 0.01339519303292036, + 0.007016956340521574, + 0.011296307668089867, + 0.008769942447543144, + 0.011387614533305168, + 0.0043777660466730595, + 0.0025082221254706383, + 0.0025923477951437235, + 0.004342709202319384, + 0.0037505703512579203, + 0.008985879831016064, + 0.0037184315733611584, + 0.0094840619713068, + 0.0035364562645554543, + 0.0028802992310374975, + 0.0033186664804816246, + 0.007674060296267271, + 0.005475969519466162, + 0.015554383397102356, + 0.00617496483027935, + 0.015580223873257637, + 0.07652623951435089, + 0.025047453120350838, + 0.0458858460187912, + 0.050669167190790176, + 0.03794822469353676, + 0.10635488480329514, + 0.06521118432283401, + 0.16629579663276672, + 0.0029006986878812313, + 0.001905654789879918, + 0.004344054963439703, + 0.0019542414229363203, + 0.0016431844560429454, + 0.005564694758504629, + 0.002085207961499691, + 0.004906147252768278, + 0.006619006395339966, + 0.004767711274325848, + 0.010428367182612419, + 0.004912371281534433, + 0.006636477541178465, + 0.006956356577575207, + 0.008885305374860764, + 0.011552072130143642, + 0.0035868233535438776, + 0.0017497885273769498, + 0.0031500908080488443, + 0.0017695014830678701, + 0.001654484891332686, + 0.0022797957062721252, + 0.0015485039912164211, + 0.00228820089250803, + 0.005043887533247471, + 0.005692809354513884, + 0.004989987704902887, + 0.003352438798174262, + 0.003921316936612129, + 0.006690233014523983, + 0.002579091116786003, + 0.006002721842378378, + 0.005583911668509245, + 0.009656498208642006, + 0.005102194845676422, + 0.007111500948667526, + 0.018747473135590553, + 0.010934052988886833, + 0.007941006682813168, + 0.0057309637777507305, + 0.0035824370570480824, + 0.0044959597289562225, + 0.0073897745460271835, + 0.00600580545142293, + 0.005130014847964048, + 0.0052800970152020454, + 0.0033520213328301907, + 0.005928912665694952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0006344689172692597 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017146660247817636 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017146660247817636, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017146660247817636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09116911888122559, + 0.055628448724746704, + 0.06309588253498077, + 0.23240070044994354, + 0.06529024243354797, + 0.08218453079462051, + 0.04590390622615814, + 0.21010306477546692, + 0.08063749969005585, + 0.0735865980386734 + ], + "credit_map": [ + 0.008514035493135452, + 0.008503919467329979, + 0.006678191479295492, + 0.01721118576824665, + 0.009713798761367798, + 0.014017695561051369, + 0.011491397395730019, + 0.015038901939988136, + 0.005339197814464569, + 0.0033557419665157795, + 0.003543838392943144, + 0.005921975243836641, + 0.00489185843616724, + 0.013205127790570259, + 0.005120603367686272, + 0.01425011083483696, + 0.004114331677556038, + 0.0033884041476994753, + 0.0041197906248271465, + 0.009602623991668224, + 0.00653971079736948, + 0.01383338961750269, + 0.008020940236747265, + 0.013476695865392685, + 0.022272448986768723, + 0.009622967801988125, + 0.017466405406594276, + 0.021884549409151077, + 0.019311562180519104, + 0.04639129340648651, + 0.030625373125076294, + 0.0648261085152626, + 0.007122684735804796, + 0.005592107307165861, + 0.012917906045913696, + 0.007343227509409189, + 0.006268886849284172, + 0.010214729234576225, + 0.006541122682392597, + 0.009289577603340149, + 0.009121138602495193, + 0.0071164402179419994, + 0.014743869192898273, + 0.0068497867323458195, + 0.007558715529739857, + 0.010316197760403156, + 0.008704881183803082, + 0.017773505300283432, + 0.008365899324417114, + 0.006519625894725323, + 0.007236461155116558, + 0.004045989830046892, + 0.006788133177906275, + 0.005654461216181517, + 0.0035616171080619097, + 0.003731717821210623, + 0.029026584699749947, + 0.02424544095993042, + 0.02699020877480507, + 0.017904236912727356, + 0.018916308879852295, + 0.028860075399279594, + 0.01735328882932663, + 0.04680692404508591, + 0.011112741194665432, + 0.012327456846833229, + 0.008710343390703201, + 0.007264446467161179, + 0.015693217515945435, + 0.010935633443295956, + 0.008894923143088818, + 0.005698738619685173, + 0.007253157906234264, + 0.010041272267699242, + 0.010652457363903522, + 0.009420914575457573, + 0.01001191046088934, + 0.009703379124403, + 0.00706787034869194, + 0.009435637854039669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0017146660247817636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot zooms in on a futuristic electronic door lock mounted on a heavy metal door. Just as the camera is about to make contact, the entire frame switches to an extreme high-speed photography style, as if time has frozen. Crafted from smooth, brushed black titanium, the lock, devoid of physical buttons and marked only by a slowly breathing, ring of eerie blue LED lights, suddenly explodes from within, triggered by an invisible force. The lock violently explodes from the center outward, sending countless gleaming metal fragments, delicate internal gears, glowing microcircuits, and delicate copper wires radiating through the air in a magnificent dance of destruction. Each swirling fragment distinctly reflects the cold white light of the surrounding environment, while tiny arcs of blue electricity flicker between the fractured components, stretching the silent explosion into a visual spectacle of technological beauty and violent impact, creating a breathtaking atmosphere of tension where the moment feels like eternity.", + "seed": 24697, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24697_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": -0.020012421428691595, + "vlm_logprob": -0.020012421428691595, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.020012421428691595, + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03188677877187729 + ], + "opposite_logprob_tokens": [ + -3.7818868160247803 + ], + "logprob_reward": -0.03188677877187729, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03188677877187729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09132857620716095, + 0.016216952353715897, + 0.012217115610837936, + 0.014275137335062027, + 0.030701937153935432, + 0.010020134039223194, + 0.769481360912323, + 0.019078893586993217, + 0.015589164569973946, + 0.021090781316161156 + ], + "credit_map": [ + 0.004830351565033197, + 0.004098635166883469, + 0.0068836817517876625, + 0.004253321792930365, + 0.0013845175271853805, + 0.03298548236489296, + 0.001476663863286376, + 0.03541592136025429, + 0.001634330372326076, + 0.0024331577587872744, + 0.002539684297516942, + 0.0030443910509347916, + 0.0008429036242887378, + 0.002189505845308304, + 0.0008286584634333849, + 0.0027043211739510298, + 0.0017026031855493784, + 0.001608425984159112, + 0.001208149129524827, + 0.0013676845701411366, + 0.0011770216515287757, + 0.002260212553665042, + 0.0009893891401588917, + 0.0019036292796954513, + 0.0034134890884160995, + 0.0016227745218202472, + 0.001999616390094161, + 0.0010240169940516353, + 0.0013367279898375273, + 0.0014712275005877018, + 0.0010238763643428683, + 0.002383408835157752, + 0.0047763376496732235, + 0.002695165341719985, + 0.00307117635384202, + 0.0018677162006497383, + 0.003021738724783063, + 0.006470124702900648, + 0.003907418809831142, + 0.0048922584392130375, + 0.001519177807494998, + 0.0009191193385049701, + 0.0009384856675751507, + 0.0009856034303084016, + 0.0006504236371256411, + 0.0016948941629379988, + 0.0014325975207611918, + 0.0018798330565914512, + 0.09049415588378906, + 0.02346884272992611, + 0.1776905059814453, + 0.043577101081609726, + 0.018442591652274132, + 0.07645493000745773, + 0.04650532081723213, + 0.2928479015827179, + 0.0015804095892235637, + 0.0008949865587055683, + 0.003910690546035767, + 0.0013533420860767365, + 0.0012793473433703184, + 0.0030833985656499863, + 0.0020509168971329927, + 0.0049258023500442505, + 0.002640306716784835, + 0.002057798905298114, + 0.0023305744398385286, + 0.0010756617411971092, + 0.0012171794660389423, + 0.002030315576121211, + 0.001869692699983716, + 0.0023676366545259953, + 0.003434209618717432, + 0.001703303074464202, + 0.002970688045024872, + 0.0013249116018414497, + 0.0017654852708801627, + 0.0032486491836607456, + 0.0035110090393573046, + 0.0031325253657996655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.03188677877187729 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205993518233299 + ], + "opposite_logprob_tokens": [ + -2.8120598793029785 + ], + "logprob_reward": -0.06205993518233299, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205993518233299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14464011788368225, + 0.050282832235097885, + 0.04911497235298157, + 0.01952248439192772, + 0.039100728929042816, + 0.03125924617052078, + 0.5304421782493591, + 0.08002612739801407, + 0.029945582151412964, + 0.025665711611509323 + ], + "credit_map": [ + 0.00850656908005476, + 0.012642267160117626, + 0.013782278634607792, + 0.011466060765087605, + 0.003519930876791477, + 0.04488131403923035, + 0.00367127638310194, + 0.046170417219400406, + 0.0032431972213089466, + 0.00781713891774416, + 0.005041827447712421, + 0.010257643647491932, + 0.004293364472687244, + 0.007736789528280497, + 0.004901452921330929, + 0.006991418544203043, + 0.01440677884966135, + 0.00250813364982605, + 0.015011399984359741, + 0.002766443882137537, + 0.002075694967061281, + 0.005846991203725338, + 0.0014859214425086975, + 0.005013610702008009, + 0.002850022865459323, + 0.002282179659232497, + 0.0018205809174105525, + 0.0018295847112312913, + 0.0022631341125816107, + 0.0028440712485462427, + 0.0021721741650253534, + 0.0034607378765940666, + 0.004442254081368446, + 0.0034492486156523228, + 0.00355179188773036, + 0.0028976532630622387, + 0.006274206563830376, + 0.004237482789903879, + 0.006369418930262327, + 0.007878671400249004, + 0.003797871293500066, + 0.004491373896598816, + 0.002449121093377471, + 0.0033272546716034412, + 0.0033650784753262997, + 0.004690015222877264, + 0.004251523409038782, + 0.004887007642537355, + 0.062294915318489075, + 0.01932096667587757, + 0.1190766841173172, + 0.02960757352411747, + 0.014422868378460407, + 0.05281288921833038, + 0.03298637643456459, + 0.1999198943376541, + 0.008165952749550343, + 0.005499749444425106, + 0.008326043374836445, + 0.006136373151093721, + 0.007085960358381271, + 0.01510266400873661, + 0.01392070297151804, + 0.01578867807984352, + 0.004409136716276407, + 0.002775125205516815, + 0.004482261836528778, + 0.002142238197848201, + 0.0029853724408894777, + 0.0036300045903772116, + 0.005357657093554735, + 0.004163786303251982, + 0.003761106403544545, + 0.0018526810454204679, + 0.0031221555545926094, + 0.002769645769149065, + 0.0035164393484592438, + 0.003266575513407588, + 0.004567591939121485, + 0.002809516154229641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.06205993518233299 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022024677600711584 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022024677600711584, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022024677600711584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09331817179918289, + 0.0729597881436348, + 0.07838515192270279, + 0.01640307530760765, + 0.025738826021552086, + 0.021282531321048737, + 0.5487164855003357, + 0.0985788106918335, + 0.021727299317717552, + 0.022889887914061546 + ], + "credit_map": [ + 0.005588472355157137, + 0.01178070344030857, + 0.01264849491417408, + 0.010527561418712139, + 0.004323774483054876, + 0.020684389397501945, + 0.004868425894528627, + 0.02289634570479393, + 0.006619886960834265, + 0.009335738606750965, + 0.005217265337705612, + 0.009171238169074059, + 0.010856807231903076, + 0.009362753480672836, + 0.011256405152380466, + 0.011139696463942528, + 0.01772323250770569, + 0.006513630505651236, + 0.012920649722218513, + 0.005545183550566435, + 0.008850489743053913, + 0.011230051517486572, + 0.006268157623708248, + 0.009333755820989609, + 0.0024011891800910234, + 0.0018683405360206962, + 0.0014581828145310283, + 0.0013902746140956879, + 0.001874924753792584, + 0.002466177800670266, + 0.001989533891901374, + 0.002954452531412244, + 0.0045335968025028706, + 0.0021218752954155207, + 0.0021419846452772617, + 0.001743042841553688, + 0.0029218397103250027, + 0.00419922498986125, + 0.003712476696819067, + 0.004364784341305494, + 0.0025003759656101465, + 0.0020417491905391216, + 0.0017764513613656163, + 0.002165976446121931, + 0.0018565634964033961, + 0.005102814175188541, + 0.003393806517124176, + 0.00244479370303452, + 0.06471893936395645, + 0.01791348122060299, + 0.12478668987751007, + 0.030632145702838898, + 0.014296283014118671, + 0.054890744388103485, + 0.033681824803352356, + 0.2077963650226593, + 0.00774613581597805, + 0.005717764608561993, + 0.007480266038328409, + 0.006579175591468811, + 0.00735453749075532, + 0.02652403712272644, + 0.013368811458349228, + 0.023808086290955544, + 0.0034579273778945208, + 0.002375120297074318, + 0.0036594136618077755, + 0.0017763470532372594, + 0.0022891145199537277, + 0.003009629435837269, + 0.0026467370335012674, + 0.0025130100548267365, + 0.003238145960494876, + 0.0019755372777581215, + 0.0029448699206113815, + 0.0024834824725985527, + 0.0028141408693045378, + 0.0028667401056736708, + 0.003259677905589342, + 0.0033072943333536386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0022024677600711584 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007189311436377466 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007189311436377466, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007189311436377466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1457255333662033, + 0.04762531816959381, + 0.039275337010622025, + 0.023426514118909836, + 0.04045022651553154, + 0.0669240728020668, + 0.3783697485923767, + 0.16569159924983978, + 0.044398754835128784, + 0.0481128953397274 + ], + "credit_map": [ + 0.009661420248448849, + 0.011103576049208641, + 0.01247453223913908, + 0.011198882944881916, + 0.005117350723594427, + 0.04514786973595619, + 0.0049757640808820724, + 0.04604613780975342, + 0.0042295451276004314, + 0.008490047417581081, + 0.004433571361005306, + 0.00832312274724245, + 0.003562653437256813, + 0.00857098214328289, + 0.0026881250087171793, + 0.007327268831431866, + 0.004642364103347063, + 0.003274636808782816, + 0.009948084130883217, + 0.0025884404312819242, + 0.0034298680257052183, + 0.0064966315403580666, + 0.0027429768815636635, + 0.0061523327603936195, + 0.0032610215712338686, + 0.003522925078868866, + 0.002408621832728386, + 0.002264967653900385, + 0.0028439394664019346, + 0.0030740704387426376, + 0.0018106502247974277, + 0.004240316804498434, + 0.004500060807913542, + 0.003410945413634181, + 0.0029546262230724096, + 0.0030170483514666557, + 0.006313508376479149, + 0.0076796989887952805, + 0.006244634743779898, + 0.006329704076051712, + 0.009313407354056835, + 0.0070717972703278065, + 0.005177013576030731, + 0.006990289781242609, + 0.007830889895558357, + 0.00805477611720562, + 0.013539167121052742, + 0.008946734480559826, + 0.042306773364543915, + 0.01792709343135357, + 0.08627825975418091, + 0.023618236184120178, + 0.010621546767652035, + 0.037129104137420654, + 0.025306284427642822, + 0.1351824551820755, + 0.010309739038348198, + 0.005369158927351236, + 0.025983473286032677, + 0.006482125259935856, + 0.00706564262509346, + 0.040401458740234375, + 0.012859788723289967, + 0.0572202131152153, + 0.004557065665721893, + 0.004551833961158991, + 0.00542013393715024, + 0.0038625127635896206, + 0.004389627370983362, + 0.005647347774356604, + 0.009613938629627228, + 0.006356296129524708, + 0.005783221684396267, + 0.0032813516445457935, + 0.008254511281847954, + 0.0036038546822965145, + 0.005406661424785852, + 0.006078837905079126, + 0.009793264791369438, + 0.00591118773445487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0007189311436377466 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031939942855387926 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.0031939942855387926, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031939942855387926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035863813012838364, + 0.02547496184706688, + 0.02911539003252983, + 0.012350980192422867, + 0.025226566940546036, + 0.1134597659111023, + 0.5385147333145142, + 0.14745783805847168, + 0.05478648468852043, + 0.017749525606632233 + ], + "credit_map": [ + 0.004436373710632324, + 0.006650437600910664, + 0.005239689722657204, + 0.00538303516805172, + 0.001927303965203464, + 0.00523883244022727, + 0.0019684843719005585, + 0.005019649863243103, + 0.002155815251171589, + 0.004697143100202084, + 0.0018588805105537176, + 0.00535521050915122, + 0.0023460621014237404, + 0.00365969305858016, + 0.0020036634523421526, + 0.003398491535335779, + 0.007983587682247162, + 0.0019029590766876936, + 0.008448508568108082, + 0.0016273021465167403, + 0.0015598548343405128, + 0.0031450598035007715, + 0.0013362803729251027, + 0.0031118374317884445, + 0.0019136543851345778, + 0.0013197994558140635, + 0.0010903903748840094, + 0.0009821983985602856, + 0.0015266237314790487, + 0.0018796788062900305, + 0.0009316680952906609, + 0.002706965897232294, + 0.0028606720734387636, + 0.002945522079244256, + 0.0023332766722887754, + 0.00243307463824749, + 0.004009442403912544, + 0.003203625325113535, + 0.00429788650944829, + 0.003143067704513669, + 0.01285631489008665, + 0.012940128333866596, + 0.012050865218043327, + 0.011780038475990295, + 0.01180198322981596, + 0.019351337105035782, + 0.01819908246397972, + 0.01448001153767109, + 0.06266039609909058, + 0.01845462992787361, + 0.1249016597867012, + 0.031056301668286324, + 0.013447651639580727, + 0.05252624303102493, + 0.03290459141135216, + 0.20256321132183075, + 0.008695738390088081, + 0.00444359565153718, + 0.02057448960840702, + 0.003934260457754135, + 0.006267847027629614, + 0.04107711464166641, + 0.008867268450558186, + 0.05359750986099243, + 0.008390625938773155, + 0.005487324669957161, + 0.013886641710996628, + 0.003275626339018345, + 0.005627321545034647, + 0.008190012536942959, + 0.004702789708971977, + 0.005226136185228825, + 0.0031281746923923492, + 0.0013312989613041282, + 0.002194739179685712, + 0.0012636397732421756, + 0.0019488142570480704, + 0.0021994065027683973, + 0.0035470242146402597, + 0.002136428141966462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0031939942855387926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sharp, dynamic swooping shot zooms in on a futuristic electronic door lock mounted on a heavy metal door. Just as the camera is about to make contact, the entire frame switches to an extreme high-speed photography style, as if time has frozen. Crafted from smooth, brushed black titanium, the lock, devoid of physical buttons and marked only by a slowly breathing, ring of eerie blue LED lights, suddenly explodes from within, triggered by an invisible force. The lock violently explodes from the center outward, sending countless gleaming metal fragments, delicate internal gears, glowing microcircuits, and delicate copper wires radiating through the air in a magnificent dance of destruction. Each swirling fragment distinctly reflects the cold white light of the surrounding environment, while tiny arcs of blue electricity flicker between the fractured components, stretching the silent explosion into a visual spectacle of technological beauty and violent impact, creating a breathtaking atmosphere of tension where the moment feels like eternity.", + "seed": 24698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_3/index_7219_seed_24698_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": -0.007828228827565908, + "vlm_logprob": -0.007828228827565908, + "vlm_yes_no_margin": 5.65, + "vlm_token_credit": -0.007828228827565908, + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013395900838077068 + ], + "opposite_logprob_tokens": [ + -4.6383957862854 + ], + "logprob_reward": -0.013395900838077068, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013395900838077068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07315012067556381, + 0.4231843948364258, + 0.2860318720340729, + 0.01745312474668026, + 0.0590406134724617, + 0.04632972180843353, + 0.04017823189496994, + 0.024691402912139893, + 0.01355715375393629, + 0.01638331264257431 + ], + "credit_map": [ + 0.015229233540594578, + 0.002897391328588128, + 0.011786788702011108, + 0.004005092661827803, + 0.0035473324824124575, + 0.017161639407277107, + 0.004423702601343393, + 0.014098933897912502, + 0.10326259583234787, + 0.01607036590576172, + 0.057441554963588715, + 0.025465788319706917, + 0.023117590695619583, + 0.1011454239487648, + 0.03438624367117882, + 0.0622948482632637, + 0.06751030683517456, + 0.010148032568395138, + 0.04142659530043602, + 0.018050292506814003, + 0.013760514557361603, + 0.06593212485313416, + 0.02548263408243656, + 0.043721385300159454, + 0.0018211210845038295, + 0.0016753844683989882, + 0.0023324214853346348, + 0.002223758026957512, + 0.001655428553931415, + 0.0021312355529516935, + 0.0018068229546770453, + 0.0038069523870944977, + 0.008362187072634697, + 0.008868790231645107, + 0.008357959799468517, + 0.009542989544570446, + 0.005416746251285076, + 0.0059591480530798435, + 0.004631343297660351, + 0.00790144968777895, + 0.00652732327580452, + 0.005879395175725222, + 0.004865681286901236, + 0.004835211206227541, + 0.008328387513756752, + 0.005697275511920452, + 0.005309595260769129, + 0.004886852111667395, + 0.0047111534513533115, + 0.002681794809177518, + 0.008955656550824642, + 0.004660325590521097, + 0.003021048614755273, + 0.0026822832878679037, + 0.0031684183049947023, + 0.010297552682459354, + 0.0019572379533201456, + 0.003702897811308503, + 0.0036915044765919447, + 0.003354014130309224, + 0.003768813330680132, + 0.0033868656028062105, + 0.0026323022320866585, + 0.0021977669093757868, + 0.0013154804473742843, + 0.0021830713376402855, + 0.0015769625315442681, + 0.0010203808778896928, + 0.0028835448902100325, + 0.0015857157995924354, + 0.0017407077830284834, + 0.0012512896209955215, + 0.002095213159918785, + 0.0020248712971806526, + 0.0019494348671287298, + 0.0016006240621209145, + 0.002286210423335433, + 0.0028000525198876858, + 0.0018536022398620844, + 0.0017733043059706688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.013395900838077068 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058347500860691 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058347500860691, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058347500860691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19930259883403778, + 0.08622387796640396, + 0.3883485198020935, + 0.01919601671397686, + 0.12908895313739777, + 0.0804586187005043, + 0.03217247128486633, + 0.038139473646879196, + 0.012698523700237274, + 0.014370903372764587 + ], + "credit_map": [ + 0.045385412871837616, + 0.0076600899919867516, + 0.032080940902233124, + 0.014705932699143887, + 0.010563761927187443, + 0.03925861790776253, + 0.017287973314523697, + 0.03235987573862076, + 0.01761670410633087, + 0.0033888265024870634, + 0.0122610442340374, + 0.006848703138530254, + 0.006034608464688063, + 0.01823897287249565, + 0.009141485206782818, + 0.012693529948592186, + 0.09626930207014084, + 0.014209134504199028, + 0.056939225643873215, + 0.023625599220395088, + 0.018308501690626144, + 0.09262854605913162, + 0.02807246893644333, + 0.058295756578445435, + 0.0019894742872565985, + 0.0019740662537515163, + 0.0022707993630319834, + 0.0030090773943811655, + 0.002670526271685958, + 0.0022609164007008076, + 0.002556620631366968, + 0.002464535878971219, + 0.010406780987977982, + 0.01594797521829605, + 0.01540020015090704, + 0.021065805107355118, + 0.01605171337723732, + 0.012860619463026524, + 0.013220354914665222, + 0.024135513231158257, + 0.0112270163372159, + 0.009928923100233078, + 0.009758356027305126, + 0.008104579523205757, + 0.011297943070530891, + 0.011064314283430576, + 0.008488720282912254, + 0.010588765144348145, + 0.0024105238262563944, + 0.002370049711316824, + 0.007822119630873203, + 0.004423861391842365, + 0.003154685953631997, + 0.0018575667636469007, + 0.0022000230383127928, + 0.007933644577860832, + 0.004033920355141163, + 0.005438076797872782, + 0.005698211491107941, + 0.004834657534956932, + 0.007582831662148237, + 0.003604272613301873, + 0.00363069586455822, + 0.0033168101217597723, + 0.0008896630024537444, + 0.002312753116711974, + 0.0010001934133470058, + 0.0010808095103129745, + 0.0033199703320860863, + 0.0014672213001176715, + 0.0011275766883045435, + 0.0015003372682258487, + 0.0017454999033361673, + 0.002032430376857519, + 0.002014939207583666, + 0.0011682084295898676, + 0.0018502094317227602, + 0.0015508237993344665, + 0.002208366524428129, + 0.0018004263984039426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02058347500860691 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013367536012083292 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013367536012083292, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013367536012083292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061893150210380554, + 0.08945176750421524, + 0.5801798701286316, + 0.04435413330793381, + 0.054018210619688034, + 0.04281222075223923, + 0.0670991986989975, + 0.01763387955725193, + 0.016739198938012123, + 0.025818385183811188 + ], + "credit_map": [ + 0.006136225536465645, + 0.0023292575497180223, + 0.00943755079060793, + 0.006545029580593109, + 0.003254394745454192, + 0.013839595019817352, + 0.0080574294552207, + 0.012293664738535881, + 0.020210912451148033, + 0.0029815665911883116, + 0.012524635531008244, + 0.0053415377624332905, + 0.004221121780574322, + 0.020913248881697655, + 0.008087954483926296, + 0.015170781873166561, + 0.1403615027666092, + 0.020866574719548225, + 0.0847189649939537, + 0.035317838191986084, + 0.027541903778910637, + 0.13697755336761475, + 0.04495822265744209, + 0.08943729102611542, + 0.004951174836605787, + 0.008120747283101082, + 0.005056608468294144, + 0.005591200664639473, + 0.004882602021098137, + 0.006573339458554983, + 0.0037013611290603876, + 0.00547709921374917, + 0.0060401917435228825, + 0.007490992080420256, + 0.007508624345064163, + 0.008656072430312634, + 0.004508481360971928, + 0.005453021731227636, + 0.005794832017272711, + 0.008565993048250675, + 0.005940438713878393, + 0.005289644468575716, + 0.004894002806395292, + 0.004149063490331173, + 0.007013465743511915, + 0.005892645567655563, + 0.0041012875735759735, + 0.005531669594347477, + 0.005393332336097956, + 0.0037884030025452375, + 0.01737123169004917, + 0.007377287372946739, + 0.004803606774657965, + 0.00359738664701581, + 0.0042424509301781654, + 0.0205254927277565, + 0.0013610238675028086, + 0.002677495591342449, + 0.0015547117218375206, + 0.0017716087168082595, + 0.004257436841726303, + 0.002904297783970833, + 0.0017800696659833193, + 0.0013272362994030118, + 0.001662283786572516, + 0.0026528320740908384, + 0.0013546125264838338, + 0.0012639854103326797, + 0.004177904687821865, + 0.002174691529944539, + 0.0017578472616150975, + 0.001695043407380581, + 0.002633861033245921, + 0.0028172035235911608, + 0.0034099912736564875, + 0.0022009825333952904, + 0.0028559600468724966, + 0.003652480198070407, + 0.0047686900943517685, + 0.0034792167134582996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.0013367536012083292 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006335158832371235 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006335158832371235, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006335158832371235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31288695335388184, + 0.060227587819099426, + 0.1386595219373703, + 0.11738414317369461, + 0.10047563165426254, + 0.12432517856359482, + 0.03815515339374542, + 0.027096936479210854, + 0.03183132782578468, + 0.04895753413438797 + ], + "credit_map": [ + 0.07509342581033707, + 0.01142878271639347, + 0.044064562767744064, + 0.02312638983130455, + 0.015519272536039352, + 0.06705373525619507, + 0.03094879351556301, + 0.04565197229385376, + 0.009909320622682571, + 0.0026478224899619818, + 0.009803036227822304, + 0.003983651287853718, + 0.003799620782956481, + 0.016284164041280746, + 0.004325747024267912, + 0.009474220685660839, + 0.033085502684116364, + 0.005377506837248802, + 0.021041441708803177, + 0.008161389268934727, + 0.006863712333142757, + 0.03029889427125454, + 0.011222319677472115, + 0.02260875329375267, + 0.011704452335834503, + 0.020820900797843933, + 0.014219775795936584, + 0.017018772661685944, + 0.014261379837989807, + 0.013245765119791031, + 0.011718355119228363, + 0.0143947284668684, + 0.013956272974610329, + 0.008987357839941978, + 0.01119427289813757, + 0.013645782135426998, + 0.011079010553658009, + 0.013487988151609898, + 0.009640899486839771, + 0.018484044820070267, + 0.016420084983110428, + 0.015982801094651222, + 0.01696818135678768, + 0.01379442773759365, + 0.016811201348900795, + 0.016082288697361946, + 0.014330404810607433, + 0.013935769908130169, + 0.005582122132182121, + 0.003218048717826605, + 0.006837537046521902, + 0.006295796949416399, + 0.0038496109191328287, + 0.0023541857954114676, + 0.0039053393993526697, + 0.006112509872764349, + 0.002202880335971713, + 0.005132246296852827, + 0.0036285333335399628, + 0.003191659227013588, + 0.003911423031240702, + 0.004237490240484476, + 0.0029122941195964813, + 0.0018804079154506326, + 0.0032246208284050226, + 0.0029489737935364246, + 0.0030646580271422863, + 0.0036554865073412657, + 0.0035101519897580147, + 0.006023861933499575, + 0.0036138957366347313, + 0.005789673887193203, + 0.005294125527143478, + 0.005301206838339567, + 0.009187177754938602, + 0.004924319218844175, + 0.005689640995115042, + 0.005998634267598391, + 0.0073883868753910065, + 0.005174038000404835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0006335158832371235 + }, + { + "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 futuristic electronic door lock mounted on a heavy metal door?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lock crafted from smooth, brushed black titanium without physical buttons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lock feature a slowly breathing ring of eerie blue LED lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lock explode from the center outward, revealing internal gears, microcircuits, and copper wires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal fragments reflect cold white light and show tiny arcs of blue electricity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045492351055145264, + 0.16955424845218658, + 0.09631398320198059, + 0.03834812343120575, + 0.13390448689460754, + 0.1263832002878189, + 0.26666098833084106, + 0.09552056342363358, + 0.011109092272818089, + 0.016712913289666176 + ], + "credit_map": [ + 0.0067760092206299305, + 0.0027475652750581503, + 0.007439910899847746, + 0.004029650241136551, + 0.002503484021872282, + 0.010348618030548096, + 0.00562225840985775, + 0.006024850532412529, + 0.041420236229896545, + 0.006425878964364529, + 0.023843050003051758, + 0.009467869065701962, + 0.008801803924143314, + 0.04010500758886337, + 0.013083528727293015, + 0.026406871154904366, + 0.022908013314008713, + 0.0037288351450115442, + 0.01429309044033289, + 0.006400002166628838, + 0.005090600810945034, + 0.02247668243944645, + 0.007346687838435173, + 0.014070075005292892, + 0.004288473632186651, + 0.008458349853754044, + 0.004639481194317341, + 0.004320263396948576, + 0.0038161631673574448, + 0.004950026981532574, + 0.0029684987384825945, + 0.004906869027763605, + 0.013998722657561302, + 0.012771246954798698, + 0.01654629409313202, + 0.015261154621839523, + 0.015077158808708191, + 0.016463320702314377, + 0.014052817597985268, + 0.02973376400768757, + 0.017503896728157997, + 0.015434572473168373, + 0.01640106737613678, + 0.013215315528213978, + 0.018410421907901764, + 0.016676845028996468, + 0.013534493744373322, + 0.015206590294837952, + 0.021844269707798958, + 0.01578202098608017, + 0.07092662155628204, + 0.03140295669436455, + 0.02114541456103325, + 0.013577320612967014, + 0.016388069838285446, + 0.07559431344270706, + 0.006809934042394161, + 0.015376108698546886, + 0.007990016601979733, + 0.008126972243189812, + 0.027409713715314865, + 0.01259816437959671, + 0.011058143340051174, + 0.006151513196527958, + 0.0008079144172370434, + 0.0017970309127122164, + 0.000966393097769469, + 0.0012352507328614593, + 0.0021880893036723137, + 0.001591606647707522, + 0.0012349081225693226, + 0.0012878982815891504, + 0.0019325207686051726, + 0.002148565137758851, + 0.0028531395364552736, + 0.0016300184652209282, + 0.0021482508163899183, + 0.00182828726246953, + 0.002057013101875782, + 0.0021151204127818346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.0031914988067001104 + } + ], + "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/0000167/group_4/index_tensor([8984, 8984])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_tensor([8984, 8984])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a9ace5e9958e55dd3c67d146158ae6fa566524cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_tensor([8984, 8984])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with a handheld camera from a slightly elevated angle, the film captures an extreme close-up of a pristine tube of lipstick in a subtle and intimate motion. The lipstick bullet, a perfect cylinder of deep velvety crimson, extends perfectly from its smooth, glossy black metal casing, reflecting a single, warm, soft light source. It stands vertically against a dark, polished, non-reflective surface, creating a stark and dramatic contrast. As the scene begins, the tip of the crimson bullet begins to shimmer under invisible heat, its matte texture giving way to a liquid sheen. Slowly and inexorably, the solid structure begins to crumble; as its shape dissolves, the lipstick gracefully slides aside, a viscous stream of liquefied crimson flowing along its sides. A heavy droplet falls, striking the surface below, forming a small, perfect circle of shimmering liquid. As the lipstick continues to melt at an accelerating pace, more liquid joins the initial droplet, causing the pool to steadily expand, its surface catching the light and rippling like liquid silk with each minute vibration of the camera. The warm color palette and focused light create an atmosphere that is captivating, sensual, and deeply intimate, capturing quiet moments of beautiful destruction.", + "seed": 4193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_8984_seed_4193_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": -0.2865902821533382, + "vlm_logprob": -0.2865902821533382, + "vlm_yes_no_margin": 2.9750000059604647, + "vlm_token_credit": -0.2865902821533382, + "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 main subject a tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012713336385786533 + ], + "opposite_logprob_tokens": [ + -5.137713432312012 + ], + "logprob_reward": -0.012713336385786533, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012713336385786533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24775318801403046, + 0.05393550544977188, + 0.03400604799389839, + 0.06381724029779434, + 0.08992790430784225, + 0.03839964419603348, + 0.04659193009138107, + 0.11945321410894394, + 0.21483823657035828, + 0.09127715975046158 + ], + "credit_map": [ + 0.035762663930654526, + 0.027903283014893532, + 0.038551852107048035, + 0.031830694526433945, + 0.02800334058701992, + 0.029230354353785515, + 0.028902631253004074, + 0.02756834588944912, + 0.011297688819468021, + 0.005491015035659075, + 0.007251647301018238, + 0.006025717128068209, + 0.004532718099653721, + 0.0046667796559631824, + 0.007326244376599789, + 0.00734369084239006, + 0.006983937695622444, + 0.003685590112581849, + 0.005644924007356167, + 0.0034342859871685505, + 0.0036763669922947884, + 0.002662542276084423, + 0.004429955966770649, + 0.003488442162051797, + 0.021587898954749107, + 0.004669054877012968, + 0.01220446452498436, + 0.005951634608209133, + 0.003002470126375556, + 0.003222008468583226, + 0.00587446428835392, + 0.0073052383959293365, + 0.011870824731886387, + 0.01074764970690012, + 0.014397292397916317, + 0.015207255259156227, + 0.007825756445527077, + 0.012356097809970379, + 0.008609415031969547, + 0.008913611061871052, + 0.005862665828317404, + 0.005205763038247824, + 0.006725715473294258, + 0.0036845472641289234, + 0.0036119408905506134, + 0.0031450586393475533, + 0.006004404742270708, + 0.004159549716860056, + 0.0064673470333218575, + 0.007147851400077343, + 0.006309122778475285, + 0.0050268578343093395, + 0.004297452978789806, + 0.007349240127950907, + 0.004853453487157822, + 0.005140600726008415, + 0.014933465048670769, + 0.011669040657579899, + 0.03207029029726982, + 0.006818113848567009, + 0.009017142467200756, + 0.006388431414961815, + 0.006604455411434174, + 0.03195226937532425, + 0.04465383663773537, + 0.021326512098312378, + 0.05612776800990105, + 0.016015157103538513, + 0.027169449254870415, + 0.011480742134153843, + 0.013435138389468193, + 0.0246296264231205, + 0.022597188130021095, + 0.007284919265657663, + 0.019270166754722595, + 0.004936403129249811, + 0.012452570721507072, + 0.00925908051431179, + 0.006492352578788996, + 0.008984467945992947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012713336385786533 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289110600948334 + ], + "opposite_logprob_tokens": [ + -1.0539110898971558 + ], + "logprob_reward": -0.4289110600948334, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289110600948334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1252070516347885, + 0.13041655719280243, + 0.07436993718147278, + 0.15261320769786835, + 0.032823268324136734, + 0.08035768568515778, + 0.0568564347922802, + 0.23333370685577393, + 0.07794725149869919, + 0.03607483580708504 + ], + "credit_map": [ + 0.021732967346906662, + 0.013612213544547558, + 0.016936857253313065, + 0.01907113939523697, + 0.010631314478814602, + 0.008066904731094837, + 0.015618976205587387, + 0.019536685198545456, + 0.03131667152047157, + 0.01121942512691021, + 0.019223080947995186, + 0.019509946927428246, + 0.009664882905781269, + 0.006754591595381498, + 0.018767433241009712, + 0.01396053098142147, + 0.018488848581910133, + 0.005879120901226997, + 0.012826585210859776, + 0.0063542635180056095, + 0.004543374292552471, + 0.004858006723225117, + 0.012134121730923653, + 0.009285615757107735, + 0.03525548800826073, + 0.012699643149971962, + 0.025472726672887802, + 0.018465114757418633, + 0.010400968603789806, + 0.01281720120459795, + 0.015591680072247982, + 0.02191038988530636, + 0.008113806135952473, + 0.0035990376491099596, + 0.00510092219337821, + 0.004694296512752771, + 0.0021151506807655096, + 0.002868254669010639, + 0.002958345226943493, + 0.0033734578173607588, + 0.015134723857045174, + 0.014789203181862831, + 0.012998799793422222, + 0.007331707514822483, + 0.004056042525917292, + 0.008895061910152435, + 0.008806271478533745, + 0.008345883339643478, + 0.011167231015861034, + 0.009407115168869495, + 0.009099291637539864, + 0.004432339686900377, + 0.0043773348443210125, + 0.005978533998131752, + 0.005760120693594217, + 0.006634471006691456, + 0.04388747364282608, + 0.013443782925605774, + 0.09681031852960587, + 0.0056810081005096436, + 0.012354649603366852, + 0.006457222625613213, + 0.005173172801733017, + 0.049526065587997437, + 0.01178562082350254, + 0.010160344652831554, + 0.01160512212663889, + 0.005745355039834976, + 0.014005010016262531, + 0.007202543318271637, + 0.008900216780602932, + 0.008543046191334724, + 0.005693341139703989, + 0.0026859943754971027, + 0.005020122975111008, + 0.002254312625154853, + 0.005055979825556278, + 0.007258254569023848, + 0.003707679221406579, + 0.004399150609970093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.4289110600948334 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961849819868803 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961849819868803, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961849819868803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0691143274307251, + 0.05260385572910309, + 0.03543558716773987, + 0.030702710151672363, + 0.06356387585401535, + 0.05159571021795273, + 0.052300553768873215, + 0.162962406873703, + 0.2574428617954254, + 0.22427816689014435 + ], + "credit_map": [ + 0.010062751360237598, + 0.007805797271430492, + 0.009154224768280983, + 0.010177062824368477, + 0.007930283434689045, + 0.008087771013379097, + 0.008630501106381416, + 0.007265927270054817, + 0.014967081137001514, + 0.0037137779872864485, + 0.008563652634620667, + 0.006034201942384243, + 0.003633599728345871, + 0.00397716136649251, + 0.006013819016516209, + 0.00570055702701211, + 0.009627814404666424, + 0.0030605837237089872, + 0.005085008684545755, + 0.0035854049492627382, + 0.0027877383399754763, + 0.002062632003799081, + 0.004757548216730356, + 0.004468853585422039, + 0.007396564353257418, + 0.003316370537504554, + 0.004783657379448414, + 0.0031440311577171087, + 0.0024311395827680826, + 0.002155028050765395, + 0.0035062141250818968, + 0.003969701938331127, + 0.007287172134965658, + 0.01031024381518364, + 0.009015374816954136, + 0.008600709028542042, + 0.005029415711760521, + 0.0075950236059725285, + 0.009097056463360786, + 0.006628869101405144, + 0.009425568394362926, + 0.010335401631891727, + 0.008568632416427135, + 0.004383016377687454, + 0.0030789070297032595, + 0.006037556566298008, + 0.004217792302370071, + 0.005548830609768629, + 0.008801729418337345, + 0.006514114793390036, + 0.012811473570764065, + 0.004102279897779226, + 0.0032535705249756575, + 0.005819128360599279, + 0.003463833360001445, + 0.007534417323768139, + 0.04573693871498108, + 0.017372624948620796, + 0.02293432131409645, + 0.014191020280122757, + 0.008890812285244465, + 0.009705145843327045, + 0.010201698169112206, + 0.03392980992794037, + 0.056268781423568726, + 0.0171205997467041, + 0.07754215598106384, + 0.020279129967093468, + 0.022769562900066376, + 0.018475929275155067, + 0.016333594918251038, + 0.028653061017394066, + 0.06080097705125809, + 0.017234476283192635, + 0.039615027606487274, + 0.01124226301908493, + 0.03289283439517021, + 0.02959401160478592, + 0.013785025104880333, + 0.019113527610898018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.005961849819868803 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080753989517689 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080753989517689, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080753989517689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05511351674795151, + 0.031879570335149765, + 0.012859394773840904, + 0.015168623998761177, + 0.06573076546192169, + 0.01691739819943905, + 0.016512256115674973, + 0.1752970665693283, + 0.15470901131629944, + 0.45581239461898804 + ], + "credit_map": [ + 0.010565804317593575, + 0.0049676415510475636, + 0.007529500871896744, + 0.008522088639438152, + 0.00534273823723197, + 0.0038079358637332916, + 0.006004826165735722, + 0.008372982032597065, + 0.010517558082938194, + 0.002240956760942936, + 0.005561522673815489, + 0.0026561515405774117, + 0.001520919264294207, + 0.0012668960262089968, + 0.004229573532938957, + 0.0038859921041876078, + 0.0028752253856509924, + 0.0011884424602612853, + 0.0018249816494062543, + 0.0010726233012974262, + 0.0012665234971791506, + 0.0008515555527992547, + 0.0021335009951144457, + 0.0016465421067550778, + 0.004172016866505146, + 0.0012315845815464854, + 0.002252128440886736, + 0.0011871259193867445, + 0.000916009652428329, + 0.0010327905183658004, + 0.0016478887991979718, + 0.0027290787547826767, + 0.007940258830785751, + 0.007592367473989725, + 0.007735136896371841, + 0.019213726744055748, + 0.005029560532420874, + 0.004762323573231697, + 0.008080577477812767, + 0.005376817192882299, + 0.002026395872235298, + 0.002732793567702174, + 0.0023925278801470995, + 0.0021997226867824793, + 0.0014808500418439507, + 0.001859166077338159, + 0.001931136823259294, + 0.0022948067635297775, + 0.0028529309201985598, + 0.0023613441735506058, + 0.002292344579473138, + 0.0015631496207788587, + 0.0008956827805377543, + 0.0025597221683710814, + 0.001683446578681469, + 0.0023036343045532703, + 0.04466508701443672, + 0.02504115365445614, + 0.029640531167387962, + 0.012894503772258759, + 0.011462396010756493, + 0.009887689724564552, + 0.008556259796023369, + 0.03314945101737976, + 0.018659096211194992, + 0.0324028879404068, + 0.015193129889667034, + 0.011076298542320728, + 0.0328994058072567, + 0.013693233020603657, + 0.01378538180142641, + 0.016999585554003716, + 0.11923524737358093, + 0.03305061161518097, + 0.08970142155885696, + 0.023648256435990334, + 0.06834599375724792, + 0.057711392641067505, + 0.022879989817738533, + 0.041239503771066666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.011080753989517689 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9742844104766846 + ], + "opposite_logprob_tokens": [ + -0.47428444027900696 + ], + "logprob_reward": -0.9742844104766846, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9742844104766846, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08133154362440109, + 0.16865621507167816, + 0.04773712903261185, + 0.054757457226514816, + 0.08458776026964188, + 0.06652715057134628, + 0.06085950508713722, + 0.21047979593276978, + 0.13553275167942047, + 0.08953066915273666 + ], + "credit_map": [ + 0.013580096885561943, + 0.008473281748592854, + 0.012104457244277, + 0.009852214716374874, + 0.005684183444827795, + 0.0067864577285945415, + 0.010082663036882877, + 0.014768201857805252, + 0.05653483420610428, + 0.012894238345324993, + 0.029215866699814796, + 0.015222692862153053, + 0.009340419434010983, + 0.006916644982993603, + 0.018154455348849297, + 0.020377077162265778, + 0.014037157408893108, + 0.0037491314578801394, + 0.008187266997992992, + 0.004510220605880022, + 0.003415799932554364, + 0.0034114625304937363, + 0.0050255367532372475, + 0.005400558467954397, + 0.013501223176717758, + 0.005821792874485254, + 0.008776637725532055, + 0.006350629962980747, + 0.004502356052398682, + 0.0036045892629772425, + 0.005538451485335827, + 0.0066617801785469055, + 0.016662687063217163, + 0.008359978906810284, + 0.012324477545917034, + 0.009685863740742207, + 0.005915776360780001, + 0.00887532439082861, + 0.008566965349018574, + 0.014196698553860188, + 0.011845560744404793, + 0.010923930443823338, + 0.010201064869761467, + 0.005169418174773455, + 0.0046320282854139805, + 0.004790849052369595, + 0.008678465150296688, + 0.010285834781825542, + 0.010918333195149899, + 0.008937356062233448, + 0.011852024123072624, + 0.0049747987650334835, + 0.004052157513797283, + 0.007863568142056465, + 0.0056209927424788475, + 0.006640278734266758, + 0.028878025710582733, + 0.007880553603172302, + 0.07901954650878906, + 0.005486727226525545, + 0.013604179956018925, + 0.00751565583050251, + 0.005416119005531073, + 0.06267900764942169, + 0.01989060267806053, + 0.013584304600954056, + 0.028128044679760933, + 0.0100614745169878, + 0.018168257549405098, + 0.016601355746388435, + 0.012288267724215984, + 0.016810448840260506, + 0.018300466239452362, + 0.007232906762510538, + 0.014829466119408607, + 0.004510675556957722, + 0.012381494976580143, + 0.016357673332095146, + 0.004980811383575201, + 0.01093718595802784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.9742844104766846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Shot with a handheld camera from a slightly elevated angle, the film captures an extreme close-up of a pristine tube of lipstick in a subtle and intimate motion. The lipstick bullet, a perfect cylinder of deep velvety crimson, extends perfectly from its smooth, glossy black metal casing, reflecting a single, warm, soft light source. It stands vertically against a dark, polished, non-reflective surface, creating a stark and dramatic contrast. As the scene begins, the tip of the crimson bullet begins to shimmer under invisible heat, its matte texture giving way to a liquid sheen. Slowly and inexorably, the solid structure begins to crumble; as its shape dissolves, the lipstick gracefully slides aside, a viscous stream of liquefied crimson flowing along its sides. A heavy droplet falls, striking the surface below, forming a small, perfect circle of shimmering liquid. As the lipstick continues to melt at an accelerating pace, more liquid joins the initial droplet, causing the pool to steadily expand, its surface catching the light and rippling like liquid silk with each minute vibration of the camera. The warm color palette and focused light create an atmosphere that is captivating, sensual, and deeply intimate, capturing quiet moments of beautiful destruction.", + "seed": 4194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_8984_seed_4194_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": -0.18788723535835744, + "vlm_logprob": -0.18788723535835744, + "vlm_yes_no_margin": 2.374999988079071, + "vlm_token_credit": -0.18788723535835744, + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1157461404800415 + ], + "opposite_logprob_tokens": [ + -2.365746021270752 + ], + "logprob_reward": -0.1157461404800415, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1157461404800415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28540530800819397, + 0.11898349225521088, + 0.05002659931778908, + 0.055684804916381836, + 0.1648775190114975, + 0.052498288452625275, + 0.09645848721265793, + 0.05269010737538338, + 0.07485412061214447, + 0.048521287739276886 + ], + "credit_map": [ + 0.038417693227529526, + 0.016940845176577568, + 0.05291379615664482, + 0.041309431195259094, + 0.0243434626609087, + 0.02982596680521965, + 0.040855392813682556, + 0.040798723697662354, + 0.019708259031176567, + 0.011223525740206242, + 0.01876848191022873, + 0.015093682333827019, + 0.015248369425535202, + 0.01287358533591032, + 0.011933060362935066, + 0.014134529046714306, + 0.01269995141774416, + 0.0027885742019861937, + 0.00875474326312542, + 0.0035053116735070944, + 0.004498940892517567, + 0.0036846951115876436, + 0.003779173130169511, + 0.01031520776450634, + 0.014637871645390987, + 0.0035743853077292442, + 0.00923151709139347, + 0.0029807838145643473, + 0.00424491660669446, + 0.006845368072390556, + 0.005161602515727282, + 0.009008364751935005, + 0.02628973312675953, + 0.008543846197426319, + 0.02537471055984497, + 0.00775957154110074, + 0.011133464984595776, + 0.02332606539130211, + 0.016833800822496414, + 0.04561634734272957, + 0.008261608891189098, + 0.0036412382032722235, + 0.008472180925309658, + 0.003702446585521102, + 0.0047288392670452595, + 0.008081674575805664, + 0.006297633517533541, + 0.009312666952610016, + 0.013294612057507038, + 0.010725111700594425, + 0.015204638242721558, + 0.008973699063062668, + 0.009534317068755627, + 0.010621355846524239, + 0.013033239170908928, + 0.015071514993906021, + 0.00901289377361536, + 0.005194265395402908, + 0.008499523624777794, + 0.0038222833536565304, + 0.005046147853136063, + 0.006339206360280514, + 0.007406048011034727, + 0.007369742263108492, + 0.011949947103857994, + 0.009187108837068081, + 0.012690926901996136, + 0.005902295932173729, + 0.006940007675439119, + 0.010192135348916054, + 0.010471857152879238, + 0.007519848179072142, + 0.006253784988075495, + 0.006638331338763237, + 0.006995889358222485, + 0.0028663824778050184, + 0.010263832286000252, + 0.005495815072208643, + 0.005504522006958723, + 0.004502731841057539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1157461404800415 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743342995643616 + ], + "opposite_logprob_tokens": [ + -0.9743342995643616 + ], + "logprob_reward": -0.4743342995643616, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743342995643616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08769958466291428, + 0.11452790349721909, + 0.16442053020000458, + 0.14125321805477142, + 0.0923909917473793, + 0.23074853420257568, + 0.0676301047205925, + 0.04368223249912262, + 0.036877311766147614, + 0.020769517868757248 + ], + "credit_map": [ + 0.011580025777220726, + 0.0059881373308598995, + 0.01368594728410244, + 0.012947186827659607, + 0.006866517011076212, + 0.011570164933800697, + 0.008218817412853241, + 0.016842789947986603, + 0.017623910680413246, + 0.0054381960071623325, + 0.01849728636443615, + 0.015858063474297523, + 0.004285898990929127, + 0.013659321703016758, + 0.010452176444232464, + 0.028713049367070198, + 0.043655041605234146, + 0.007378664333373308, + 0.02854795567691326, + 0.012507270090281963, + 0.012143689207732677, + 0.016004696488380432, + 0.012548706494271755, + 0.03163451701402664, + 0.0371435284614563, + 0.010274833999574184, + 0.024614855647087097, + 0.007176970597356558, + 0.00941687636077404, + 0.015007853507995605, + 0.012962475419044495, + 0.02465582825243473, + 0.015544629655778408, + 0.006328159011900425, + 0.015284620225429535, + 0.0055924030020833015, + 0.007018993142992258, + 0.012925789691507816, + 0.009735370054841042, + 0.019961031153798103, + 0.027562499046325684, + 0.019008498638868332, + 0.03617328777909279, + 0.016747431829571724, + 0.033871397376060486, + 0.02911120094358921, + 0.03993145748972893, + 0.028342757374048233, + 0.010173591785132885, + 0.0059255631640553474, + 0.012001197785139084, + 0.005217639729380608, + 0.005065795499831438, + 0.008245664648711681, + 0.011404737830162048, + 0.009595915675163269, + 0.008072169497609138, + 0.004184233024716377, + 0.006591104436665773, + 0.0037008021026849747, + 0.003258685115724802, + 0.00468518678098917, + 0.0060862149111926556, + 0.007103838492184877, + 0.0047456747852265835, + 0.003590000094845891, + 0.005846728105098009, + 0.002195600187405944, + 0.0034582095686346292, + 0.006513527128845453, + 0.004848345648497343, + 0.0056792269460856915, + 0.0038925004191696644, + 0.0019416717113927007, + 0.0030923704616725445, + 0.0012563393684104085, + 0.002398786833509803, + 0.003863733494654298, + 0.001869895844720304, + 0.0024542193859815598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.4743342995643616 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561369687318802 + ], + "opposite_logprob_tokens": [ + -4.387561321258545 + ], + "logprob_reward": -0.012561369687318802, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561369687318802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16054800152778625, + 0.04869404435157776, + 0.031021298840641975, + 0.03806964308023453, + 0.09359202533960342, + 0.1320788562297821, + 0.10406764596700668, + 0.13779239356517792, + 0.1595175415277481, + 0.0946185290813446 + ], + "credit_map": [ + 0.01814422197639942, + 0.01225174218416214, + 0.029286671429872513, + 0.03940965235233307, + 0.010533088818192482, + 0.013277839869260788, + 0.019709810614585876, + 0.01793498918414116, + 0.006602267734706402, + 0.003403792856261134, + 0.007389948703348637, + 0.008860079571604729, + 0.004752520006150007, + 0.005335967522114515, + 0.005289795808494091, + 0.0070596737787127495, + 0.004817736800760031, + 0.0024796260986477137, + 0.004456918220967054, + 0.0025796943809837103, + 0.004512343555688858, + 0.003815258154645562, + 0.0033436294179409742, + 0.005016094073653221, + 0.005582637619227171, + 0.0029154557269066572, + 0.006259546149522066, + 0.003948122262954712, + 0.0034887471701949835, + 0.005235016345977783, + 0.004464672412723303, + 0.006175450049340725, + 0.012122180312871933, + 0.0066446601413190365, + 0.01633952558040619, + 0.006537444889545441, + 0.008680676110088825, + 0.013109471648931503, + 0.011432552710175514, + 0.018725527450442314, + 0.017475465312600136, + 0.005739815533161163, + 0.02393225021660328, + 0.005857578944414854, + 0.017866887152194977, + 0.026783086359500885, + 0.009992525912821293, + 0.024431247264146805, + 0.014737436547875404, + 0.011483002454042435, + 0.014341234229505062, + 0.011044268496334553, + 0.009184698574244976, + 0.0119598638266325, + 0.015238475054502487, + 0.01607867330312729, + 0.02658979967236519, + 0.012971301563084126, + 0.019958144053816795, + 0.015299622900784016, + 0.013464105315506458, + 0.015551737509667873, + 0.0206533782184124, + 0.013304315507411957, + 0.027310997247695923, + 0.01850883662700653, + 0.02509903348982334, + 0.009789183735847473, + 0.01144226174801588, + 0.025798728689551353, + 0.022121816873550415, + 0.01944669522345066, + 0.01728898659348488, + 0.00922420248389244, + 0.01475345529615879, + 0.006356917321681976, + 0.015842946246266365, + 0.011675546877086163, + 0.0126016391441226, + 0.006874836515635252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012561369687318802 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329867146909237 + ], + "opposite_logprob_tokens": [ + -3.773298740386963 + ], + "logprob_reward": -0.02329867146909237, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329867146909237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13462994992733002, + 0.04687804728746414, + 0.03312106430530548, + 0.06329638510942459, + 0.12296118587255478, + 0.0589296780526638, + 0.10277853161096573, + 0.16764257848262787, + 0.177200585603714, + 0.0925620049238205 + ], + "credit_map": [ + 0.013093952089548111, + 0.010150065645575523, + 0.022256910800933838, + 0.030606534332036972, + 0.008120262995362282, + 0.013475645333528519, + 0.01454453356564045, + 0.022382041439414024, + 0.007841971702873707, + 0.0029504175763577223, + 0.007420050445944071, + 0.004822513088583946, + 0.004591646138578653, + 0.005773826036602259, + 0.0044019450433552265, + 0.009075678884983063, + 0.0060859741643071175, + 0.0020132993813604116, + 0.005352183245122433, + 0.0034747282043099403, + 0.002592360367998481, + 0.003343330230563879, + 0.0037873038090765476, + 0.006471885833889246, + 0.010277270339429379, + 0.006412283051759005, + 0.010406353510916233, + 0.005779608152806759, + 0.005814347416162491, + 0.0069258264265954494, + 0.006729079410433769, + 0.01095161959528923, + 0.01579931005835533, + 0.008846638724207878, + 0.018897157162427902, + 0.0067084054462611675, + 0.010467924177646637, + 0.01980678178369999, + 0.01163043174892664, + 0.03080453909933567, + 0.009873302653431892, + 0.004282664507627487, + 0.009512003511190414, + 0.004339747596532106, + 0.004943439271301031, + 0.00798529852181673, + 0.009066357277333736, + 0.008926867507398129, + 0.016286592930555344, + 0.01138751208782196, + 0.015405473299324512, + 0.007827523164451122, + 0.009621861390769482, + 0.01481782365590334, + 0.014991319738328457, + 0.012440423481166363, + 0.02603241614997387, + 0.020229024812579155, + 0.024741392582654953, + 0.010204780846834183, + 0.015282281674444675, + 0.021549716591835022, + 0.022228389978408813, + 0.027374571189284325, + 0.0247709508985281, + 0.01982489973306656, + 0.030971331521868706, + 0.012676003389060497, + 0.014678481966257095, + 0.024990862235426903, + 0.02700258046388626, + 0.022285467013716698, + 0.013450687751173973, + 0.010492255911231041, + 0.01510620303452015, + 0.007333609275519848, + 0.012308390811085701, + 0.011801546439528465, + 0.01234997995197773, + 0.00971932802349329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02329867146909237 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134956955909729 + ], + "opposite_logprob_tokens": [ + -1.3134956359863281 + ], + "logprob_reward": -0.3134956955909729, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134956955909729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17155984044075012, + 0.07159550487995148, + 0.15979453921318054, + 0.04160439968109131, + 0.03170192986726761, + 0.0767493024468422, + 0.10567931830883026, + 0.13286937773227692, + 0.1574878841638565, + 0.050957899540662766 + ], + "credit_map": [ + 0.03899884968996048, + 0.010274996981024742, + 0.026738567277789116, + 0.02430422604084015, + 0.010457279160618782, + 0.015951843932271004, + 0.0168081633746624, + 0.028025930747389793, + 0.014959794469177723, + 0.004300624597817659, + 0.010292105376720428, + 0.013641130179166794, + 0.004925252869725227, + 0.006357905920594931, + 0.00741469906643033, + 0.009704000316560268, + 0.03156951069831848, + 0.008562860079109669, + 0.03212249279022217, + 0.011271684430539608, + 0.011217941530048847, + 0.017359791323542595, + 0.013637552037835121, + 0.034052714705467224, + 0.006180556025356054, + 0.003810731228441, + 0.006328808143734932, + 0.004099180921912193, + 0.003891291096806526, + 0.004743370693176985, + 0.004632207099348307, + 0.00791825633496046, + 0.0037369695492088795, + 0.002349971793591976, + 0.0043468778021633625, + 0.0024597428273409605, + 0.0038804670330137014, + 0.0051726084202528, + 0.00457082549110055, + 0.005184471607208252, + 0.012472566217184067, + 0.005124205723404884, + 0.012071295641362667, + 0.004248542245477438, + 0.007000522688031197, + 0.01043968740850687, + 0.006450622342526913, + 0.01894187368452549, + 0.0192521084100008, + 0.013358972035348415, + 0.015107812359929085, + 0.009099043905735016, + 0.00806098897010088, + 0.012600023299455643, + 0.015076139941811562, + 0.013124238699674606, + 0.018758267164230347, + 0.014322217553853989, + 0.01977112703025341, + 0.008467435836791992, + 0.013587222434580326, + 0.01960316114127636, + 0.01836560294032097, + 0.0199943408370018, + 0.027547109872102737, + 0.022300096228718758, + 0.02746601589024067, + 0.012082166038453579, + 0.01062399335205555, + 0.02405630238354206, + 0.020244281738996506, + 0.013167941942811012, + 0.0077894944697618484, + 0.006621557753533125, + 0.008008366450667381, + 0.003781410399824381, + 0.006234703119844198, + 0.007156828418374062, + 0.006486088968813419, + 0.00487945694476366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.3134956955909729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera from a slightly elevated angle, the film captures an extreme close-up of a pristine tube of lipstick in a subtle and intimate motion. The lipstick bullet, a perfect cylinder of deep velvety crimson, extends perfectly from its smooth, glossy black metal casing, reflecting a single, warm, soft light source. It stands vertically against a dark, polished, non-reflective surface, creating a stark and dramatic contrast. As the scene begins, the tip of the crimson bullet begins to shimmer under invisible heat, its matte texture giving way to a liquid sheen. Slowly and inexorably, the solid structure begins to crumble; as its shape dissolves, the lipstick gracefully slides aside, a viscous stream of liquefied crimson flowing along its sides. A heavy droplet falls, striking the surface below, forming a small, perfect circle of shimmering liquid. As the lipstick continues to melt at an accelerating pace, more liquid joins the initial droplet, causing the pool to steadily expand, its surface catching the light and rippling like liquid silk with each minute vibration of the camera. The warm color palette and focused light create an atmosphere that is captivating, sensual, and deeply intimate, capturing quiet moments of beautiful destruction.", + "seed": 4195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_8984_seed_4195_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": -0.05084835318848491, + "vlm_logprob": -0.05084835318848491, + "vlm_yes_no_margin": 3.700000023841858, + "vlm_token_credit": -0.05084835318848491, + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013910486362874508 + ], + "opposite_logprob_tokens": [ + -4.638910293579102 + ], + "logprob_reward": -0.013910486362874508, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013910486362874508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4524955153465271, + 0.1949792206287384, + 0.05249273404479027, + 0.05433766916394234, + 0.04679417237639427, + 0.06797658652067184, + 0.021837767213582993, + 0.03959829732775688, + 0.05195302516222, + 0.017535032704472542 + ], + "credit_map": [ + 0.03592151775956154, + 0.0076625170186161995, + 0.14886844158172607, + 0.009209450334310532, + 0.014504017308354378, + 0.15575888752937317, + 0.030069401487708092, + 0.05050129443407059, + 0.016712330281734467, + 0.00393559318035841, + 0.04491426795721054, + 0.007132863160222769, + 0.008815409615635872, + 0.07604095339775085, + 0.017740868031978607, + 0.019686931744217873, + 0.007937854155898094, + 0.001854299334809184, + 0.008260955102741718, + 0.0016085923416540027, + 0.003577267285436392, + 0.011674309149384499, + 0.006222227122634649, + 0.011357231996953487, + 0.008393581956624985, + 0.0034696143120527267, + 0.010802424512803555, + 0.0019752494990825653, + 0.006614249665290117, + 0.011623210273683071, + 0.005979894194751978, + 0.005479445680975914, + 0.008912669494748116, + 0.0022567224223166704, + 0.008970123715698719, + 0.0017424053512513638, + 0.004782116040587425, + 0.010114460252225399, + 0.004370606038719416, + 0.005645068362355232, + 0.009342514909803867, + 0.0032030907459557056, + 0.014973037876188755, + 0.004383069928735495, + 0.006417310796678066, + 0.014782062731683254, + 0.008089513517916203, + 0.006785988342016935, + 0.0035801378544420004, + 0.001374675310216844, + 0.00363862793892622, + 0.0016805357299745083, + 0.0027498090639710426, + 0.004256675951182842, + 0.0019687304738909006, + 0.0025885740760713816, + 0.011358697898685932, + 0.0014731144765391946, + 0.009118744172155857, + 0.0026223708409816027, + 0.0023831031285226345, + 0.005368734709918499, + 0.002104892861098051, + 0.005168643780052662, + 0.011093721725046635, + 0.003388307522982359, + 0.0107046477496624, + 0.0052855778485536575, + 0.005222134292125702, + 0.0067038931883871555, + 0.0032255847472697496, + 0.006329160183668137, + 0.0026734014973044395, + 0.001432298799045384, + 0.0024036613758653402, + 0.002043855143710971, + 0.0020755943842232227, + 0.002239945810288191, + 0.0029592288192361593, + 0.0017070473404601216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.013910486362874508 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899021357297897 + ], + "opposite_logprob_tokens": [ + -2.5789902210235596 + ], + "logprob_reward": -0.07899021357297897, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899021357297897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04821312800049782, + 0.5423079133033752, + 0.011533276177942753, + 0.0326344259083271, + 0.0464874729514122, + 0.03322986140847206, + 0.04873039945960045, + 0.10554937273263931, + 0.05813180282711983, + 0.07318226993083954 + ], + "credit_map": [ + 0.004541691858321428, + 0.001381638809107244, + 0.012126573361456394, + 0.002010714029893279, + 0.0034104869700968266, + 0.016195857897400856, + 0.004258385859429836, + 0.004287785850465298, + 0.05167454108595848, + 0.009203868918120861, + 0.12821020185947418, + 0.011377986520528793, + 0.022861646488308907, + 0.22188302874565125, + 0.04548849165439606, + 0.05160818248987198, + 0.0020704360213130713, + 0.0007913123699836433, + 0.0019428026862442493, + 0.0008593078819103539, + 0.001283603603951633, + 0.001851600711233914, + 0.0012082236353307962, + 0.0015259901992976665, + 0.004808830097317696, + 0.0023303222842514515, + 0.007410687860101461, + 0.0012463873717933893, + 0.0042503043077886105, + 0.006243202369660139, + 0.003392548765987158, + 0.0029521475080400705, + 0.009319108910858631, + 0.0018465231405571103, + 0.007269881200045347, + 0.0016807200154289603, + 0.0037532756105065346, + 0.011699088849127293, + 0.0050020283088088036, + 0.005916849710047245, + 0.005593155510723591, + 0.0017995352391153574, + 0.004614020697772503, + 0.002741699106991291, + 0.0034798283595591784, + 0.006307458505034447, + 0.0041662827134132385, + 0.004527886398136616, + 0.005016976036131382, + 0.0026320971082895994, + 0.00809311494231224, + 0.0034097800962626934, + 0.006598119158297777, + 0.006599433720111847, + 0.006525710690766573, + 0.009855173528194427, + 0.01840382255613804, + 0.006012683734297752, + 0.02486785128712654, + 0.01209838967770338, + 0.010268607176840305, + 0.015903722494840622, + 0.008148517459630966, + 0.009845790453255177, + 0.011818276718258858, + 0.003847691463306546, + 0.0116353090852499, + 0.005421703681349754, + 0.005588985979557037, + 0.00799193698912859, + 0.0046424465253949165, + 0.007185458205640316, + 0.01459368597716093, + 0.004867002833634615, + 0.013252380304038525, + 0.004266313277184963, + 0.005730590783059597, + 0.011858955956995487, + 0.005887830629944801, + 0.012725519016385078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.07899021357297897 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964220035821199 + ], + "opposite_logprob_tokens": [ + -5.130964279174805 + ], + "logprob_reward": -0.005964220035821199, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964220035821199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2193957418203354, + 0.3325576186180115, + 0.025652563199400902, + 0.029651746153831482, + 0.047698888927698135, + 0.04456113651394844, + 0.07019016146659851, + 0.059803690761327744, + 0.11656033247709274, + 0.05392806977033615 + ], + "credit_map": [ + 0.01637549139559269, + 0.003349239705130458, + 0.0686907097697258, + 0.004234431777149439, + 0.008694354444742203, + 0.08195440471172333, + 0.016057347878813744, + 0.020039767026901245, + 0.03132397308945656, + 0.005592069588601589, + 0.08012330532073975, + 0.00720483111217618, + 0.013955166563391685, + 0.1360369175672531, + 0.02768791653215885, + 0.030633430927991867, + 0.004275997169315815, + 0.0008707016240805387, + 0.003687290009111166, + 0.0008782631484791636, + 0.001519454992376268, + 0.00518526928499341, + 0.0030928230844438076, + 0.006142762955278158, + 0.004274958744645119, + 0.0015801070258021355, + 0.004748811479657888, + 0.0010059585329145193, + 0.00307404063642025, + 0.007000685669481754, + 0.0034501859918236732, + 0.004516999237239361, + 0.010247091762721539, + 0.0023498276714235544, + 0.008536791428923607, + 0.002146585611626506, + 0.004145425278693438, + 0.011136314831674099, + 0.004580314736813307, + 0.004556537605822086, + 0.007182370871305466, + 0.0021503905300050974, + 0.006432876456528902, + 0.0023825461976230145, + 0.004804231692105532, + 0.010455555282533169, + 0.0062436992302536964, + 0.00490946602076292, + 0.006184533704072237, + 0.002507166936993599, + 0.012721942737698555, + 0.002707344014197588, + 0.006958805490285158, + 0.012390217743813992, + 0.009290944784879684, + 0.017429206520318985, + 0.00997531320899725, + 0.003164644120261073, + 0.011191434226930141, + 0.006092786323279142, + 0.005950124468654394, + 0.011830361559987068, + 0.005685975309461355, + 0.005913048982620239, + 0.022515516728162766, + 0.004571861587464809, + 0.02692371979355812, + 0.005776556208729744, + 0.007676031440496445, + 0.024788327515125275, + 0.004930953960865736, + 0.019377367570996284, + 0.008704188279807568, + 0.0037264872808009386, + 0.006167867686599493, + 0.004285996314138174, + 0.007428467273712158, + 0.007818454876542091, + 0.006380660925060511, + 0.009415945969522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005964220035821199 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545594945549965 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545594945549965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545594945549965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0794379860162735, + 0.41641050577163696, + 0.03299340233206749, + 0.020965546369552612, + 0.02185812219977379, + 0.024787306785583496, + 0.05195269361138344, + 0.05417471379041672, + 0.20680195093154907, + 0.09061765670776367 + ], + "credit_map": [ + 0.006814360152930021, + 0.0016965481918305159, + 0.027577165514230728, + 0.0025813262909650803, + 0.0030527478083968163, + 0.02716570906341076, + 0.005079977214336395, + 0.0054701645858585835, + 0.03877949342131615, + 0.006600371096283197, + 0.10079912841320038, + 0.008661536499857903, + 0.017583569511771202, + 0.17058975994586945, + 0.034709759056568146, + 0.03868695721030235, + 0.005273930262774229, + 0.0012768687447533011, + 0.005351032130420208, + 0.0011238829465582967, + 0.002455379581078887, + 0.006668641697615385, + 0.003933969419449568, + 0.006909704301506281, + 0.003296076785773039, + 0.001560310018248856, + 0.0036526722833514214, + 0.000878084683790803, + 0.0029124077409505844, + 0.004099639132618904, + 0.002256580162793398, + 0.0023097798693925142, + 0.0040207793936133385, + 0.0011903728591278195, + 0.004766331519931555, + 0.001036310801282525, + 0.0015058425487950444, + 0.004868005868047476, + 0.0018797506345435977, + 0.0025907333474606276, + 0.003760390914976597, + 0.0011953107314184308, + 0.004730356391519308, + 0.000977852731011808, + 0.002808074001222849, + 0.005718617234379053, + 0.002463336568325758, + 0.0031333742663264275, + 0.0062430440448224545, + 0.0021965401247143745, + 0.00991453230381012, + 0.0018985201604664326, + 0.004701391328126192, + 0.008348356932401657, + 0.0063074673525989056, + 0.012342850677669048, + 0.013441151939332485, + 0.002442690310999751, + 0.011478523723781109, + 0.004888004623353481, + 0.004165157210081816, + 0.00880752969533205, + 0.003175257006660104, + 0.005776406731456518, + 0.03887642174959183, + 0.007757149171084166, + 0.027126288041472435, + 0.007942241616547108, + 0.011831257492303848, + 0.04827604815363884, + 0.008880526758730412, + 0.05611206218600273, + 0.015895921736955643, + 0.005263407714664936, + 0.011410748586058617, + 0.006680311635136604, + 0.011802010238170624, + 0.013734743930399418, + 0.009738673456013203, + 0.016091851517558098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.012545594945549965 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428312510251999 + ], + "opposite_logprob_tokens": [ + -2.017831325531006 + ], + "logprob_reward": -0.1428312510251999, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428312510251999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10202255100011826, + 0.11230550706386566, + 0.041251037269830704, + 0.10547291487455368, + 0.21808937191963196, + 0.10642387717962265, + 0.19471430778503418, + 0.06365171074867249, + 0.03920596092939377, + 0.016862820833921432 + ], + "credit_map": [ + 0.008224347606301308, + 0.002250135876238346, + 0.034336693584918976, + 0.002977612428367138, + 0.0048389011062681675, + 0.03483259677886963, + 0.00744423596188426, + 0.00711801927536726, + 0.010943177156150341, + 0.0019203892443329096, + 0.02942211553454399, + 0.0024547353386878967, + 0.004475730005651712, + 0.04450588300824165, + 0.009044289588928223, + 0.009539184160530567, + 0.006651428062468767, + 0.0020764509681612253, + 0.005958300083875656, + 0.0018939054571092129, + 0.003412191988900304, + 0.007136557716876268, + 0.004881821572780609, + 0.009240375831723213, + 0.016376616433262825, + 0.005768624600023031, + 0.024695545434951782, + 0.0038110846653580666, + 0.010770071297883987, + 0.023023083806037903, + 0.01037689484655857, + 0.01065098401159048, + 0.04751894995570183, + 0.008011581376194954, + 0.04706084355711937, + 0.00743497209623456, + 0.01785077340900898, + 0.047353994101285934, + 0.020450374111533165, + 0.022407859563827515, + 0.013997641392052174, + 0.004036964848637581, + 0.0263429656624794, + 0.004223587457090616, + 0.008674927055835724, + 0.02939067967236042, + 0.011628220789134502, + 0.008128881454467773, + 0.025109753012657166, + 0.009172328747808933, + 0.022684646770358086, + 0.010771460831165314, + 0.020153844729065895, + 0.03562473878264427, + 0.024978015571832657, + 0.046219516545534134, + 0.017113201320171356, + 0.002446286380290985, + 0.014671331271529198, + 0.003446468384936452, + 0.004268229939043522, + 0.008241468109190464, + 0.0037459200248122215, + 0.009718803688883781, + 0.007286520674824715, + 0.0028541486244648695, + 0.006651108153164387, + 0.004203916061669588, + 0.0031857856083661318, + 0.005193887744098902, + 0.003536180593073368, + 0.006294408347457647, + 0.002910717623308301, + 0.0011558980913832784, + 0.002439602743834257, + 0.0012361587723717093, + 0.0019288535695523024, + 0.002253642538562417, + 0.0018223830265924335, + 0.0031155634205788374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.1428312510251999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a handheld camera from a slightly elevated angle, the film captures an extreme close-up of a pristine tube of lipstick in a subtle and intimate motion. The lipstick bullet, a perfect cylinder of deep velvety crimson, extends perfectly from its smooth, glossy black metal casing, reflecting a single, warm, soft light source. It stands vertically against a dark, polished, non-reflective surface, creating a stark and dramatic contrast. As the scene begins, the tip of the crimson bullet begins to shimmer under invisible heat, its matte texture giving way to a liquid sheen. Slowly and inexorably, the solid structure begins to crumble; as its shape dissolves, the lipstick gracefully slides aside, a viscous stream of liquefied crimson flowing along its sides. A heavy droplet falls, striking the surface below, forming a small, perfect circle of shimmering liquid. As the lipstick continues to melt at an accelerating pace, more liquid joins the initial droplet, causing the pool to steadily expand, its surface catching the light and rippling like liquid silk with each minute vibration of the camera. The warm color palette and focused light create an atmosphere that is captivating, sensual, and deeply intimate, capturing quiet moments of beautiful destruction.", + "seed": 4196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_4/index_8984_seed_4196_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": -0.27430653627961876, + "vlm_logprob": -0.27430653627961876, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.27430653627961876, + "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 main subject a tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02825368009507656 + ], + "opposite_logprob_tokens": [ + -3.9032537937164307 + ], + "logprob_reward": -0.02825368009507656, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02825368009507656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2872439920902252, + 0.06916939467191696, + 0.023618565872311592, + 0.060350190848112106, + 0.0567668192088604, + 0.24430103600025177, + 0.05727580934762955, + 0.048239100724458694, + 0.0227532759308815, + 0.13028179109096527 + ], + "credit_map": [ + 0.047388724982738495, + 0.02366381138563156, + 0.06874710321426392, + 0.03166411444544792, + 0.020601116120815277, + 0.028661027550697327, + 0.03715845197439194, + 0.02935962751507759, + 0.013459539040923119, + 0.004511402919888496, + 0.009983504191040993, + 0.0061864666640758514, + 0.007271719630807638, + 0.013180487789213657, + 0.0064176470041275024, + 0.008158622309565544, + 0.0037034584674984217, + 0.0020614881068468094, + 0.003940360154956579, + 0.0020196351688355207, + 0.003324259305372834, + 0.002937142737209797, + 0.0028459192253649235, + 0.002786303171887994, + 0.009633754380047321, + 0.009113898500800133, + 0.007307013496756554, + 0.0079980893060565, + 0.006056133192032576, + 0.007182621397078037, + 0.0055173058062791824, + 0.007541375700384378, + 0.006599492859095335, + 0.009301098994910717, + 0.012489445507526398, + 0.005649687722325325, + 0.004545284900814295, + 0.007413081359118223, + 0.005607147701084614, + 0.005161583889275789, + 0.015869613736867905, + 0.08966582268476486, + 0.08629179000854492, + 0.014378727413713932, + 0.006115719210356474, + 0.019307749345898628, + 0.007368701044470072, + 0.005302901845425367, + 0.006564686074852943, + 0.013574409298598766, + 0.011828751303255558, + 0.004847823176532984, + 0.005328908562660217, + 0.0035421273205429316, + 0.006207023281604052, + 0.005382084753364325, + 0.007893458008766174, + 0.004272866528481245, + 0.010953376069664955, + 0.006723733618855476, + 0.004165223333984613, + 0.0037108492106199265, + 0.0061857509426772594, + 0.0043338448740541935, + 0.002722208620980382, + 0.0024944813922047615, + 0.00316628604196012, + 0.0023440250661224127, + 0.003045592922717333, + 0.003155274549499154, + 0.0029817824251949787, + 0.0028436249122023582, + 0.022664839401841164, + 0.03351322188973427, + 0.040800731629133224, + 0.00900704599916935, + 0.006949570961296558, + 0.005482274107635021, + 0.005929028149694204, + 0.005935079883784056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.02825368009507656 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328206062316895 + ], + "opposite_logprob_tokens": [ + -0.7578206062316895 + ], + "logprob_reward": -0.6328206062316895, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328206062316895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08889291435480118, + 0.0769595056772232, + 0.02453579381108284, + 0.03031162917613983, + 0.05576061084866524, + 0.542137861251831, + 0.05278922617435455, + 0.028446968644857407, + 0.04500337317585945, + 0.05516209453344345 + ], + "credit_map": [ + 0.010126058012247086, + 0.01442412193864584, + 0.013938568532466888, + 0.011708126403391361, + 0.012309922836720943, + 0.008043871261179447, + 0.009686996228992939, + 0.00865525845438242, + 0.018612144514918327, + 0.004943180363625288, + 0.012304790318012238, + 0.008569100871682167, + 0.007998189888894558, + 0.009421667084097862, + 0.009179930202662945, + 0.005930513143539429, + 0.004412380512803793, + 0.0020449613220989704, + 0.003549849381670356, + 0.0021252166479825974, + 0.003952814266085625, + 0.0029502015095204115, + 0.0026745276991277933, + 0.002825844567269087, + 0.004544590599834919, + 0.0034538835752755404, + 0.0041432795114815235, + 0.00379777979105711, + 0.0037870542146265507, + 0.004011723678559065, + 0.003315506037324667, + 0.0032578168902546167, + 0.008809002116322517, + 0.008065355941653252, + 0.008273689076304436, + 0.006407178472727537, + 0.0038200642447918653, + 0.007116011809557676, + 0.008151679299771786, + 0.005117637105286121, + 0.04034028574824333, + 0.2179601639509201, + 0.18911877274513245, + 0.025107989087700844, + 0.011329663917422295, + 0.03957131877541542, + 0.010554364882409573, + 0.00815536081790924, + 0.005786416120827198, + 0.0096875736489892, + 0.009821174666285515, + 0.005852791480720043, + 0.005332297179847956, + 0.0029138654936105013, + 0.006164142861962318, + 0.007230969611555338, + 0.003382984781637788, + 0.0031115070451050997, + 0.005722421687096357, + 0.003074310254305601, + 0.0026459498330950737, + 0.0031286957673728466, + 0.003594696754589677, + 0.003786405548453331, + 0.011653888039290905, + 0.0028304990846663713, + 0.014380188658833504, + 0.0045069982297718525, + 0.0025813013780862093, + 0.003115488449111581, + 0.0025449825916439295, + 0.0033900276757776737, + 0.006530431564897299, + 0.02003302052617073, + 0.010346794500946999, + 0.005143476650118828, + 0.003704397240653634, + 0.0025442149490118027, + 0.00387688004411757, + 0.0029828862752765417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.6328206062316895 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965523421764374 + ], + "opposite_logprob_tokens": [ + -5.130965709686279 + ], + "logprob_reward": -0.005965523421764374, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965523421764374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09109361469745636, + 0.0690632089972496, + 0.01941682957112789, + 0.07163802534341812, + 0.10876429080963135, + 0.06419388949871063, + 0.19672223925590515, + 0.18708810210227966, + 0.08050306886434555, + 0.11151677370071411 + ], + "credit_map": [ + 0.015508840791881084, + 0.009123780764639378, + 0.018269628286361694, + 0.012813304550945759, + 0.008333726786077023, + 0.009587811306118965, + 0.008217843249440193, + 0.009238682687282562, + 0.017401637509465218, + 0.0041940053924918175, + 0.011804725043475628, + 0.007466452661901712, + 0.007561934180557728, + 0.007040389347821474, + 0.00709000276401639, + 0.00650406489148736, + 0.0027221678756177425, + 0.0014741378836333752, + 0.0036443653516471386, + 0.001882963813841343, + 0.0021475700195878744, + 0.00266718165948987, + 0.0018230276182293892, + 0.003055415814742446, + 0.010463005863130093, + 0.008978556841611862, + 0.010692189447581768, + 0.008561206050217152, + 0.007952854037284851, + 0.008948657661676407, + 0.00849470216780901, + 0.007546853274106979, + 0.014595402404665947, + 0.013374228030443192, + 0.016955653205513954, + 0.01429247297346592, + 0.008758039213716984, + 0.018436184152960777, + 0.014837433584034443, + 0.007514879107475281, + 0.007443853188306093, + 0.017418330535292625, + 0.010267731733620167, + 0.008422262966632843, + 0.004900685045868158, + 0.0068696788512170315, + 0.005649231839925051, + 0.0032221092842519283, + 0.01804223284125328, + 0.06499035656452179, + 0.05483284592628479, + 0.013004949316382408, + 0.010425342246890068, + 0.007153823971748352, + 0.014747906476259232, + 0.013524793088436127, + 0.026476455852389336, + 0.018980437889695168, + 0.040440987795591354, + 0.01984631083905697, + 0.01815635710954666, + 0.01651979796588421, + 0.02982117421925068, + 0.016846593469381332, + 0.014954191632568836, + 0.008007263764739037, + 0.0183038879185915, + 0.008390499278903008, + 0.008237529546022415, + 0.00970507599413395, + 0.005496126599609852, + 0.007408483419567347, + 0.018033813685178757, + 0.037403881549835205, + 0.010203853249549866, + 0.013438262045383453, + 0.007891532965004444, + 0.008528758771717548, + 0.008300880901515484, + 0.007715793792158365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.005965523421764374 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081461794674397 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081461794674397, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081461794674397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07414653152227402, + 0.0339476577937603, + 0.0219100508838892, + 0.019667690619826317, + 0.043184466660022736, + 0.035891901701688766, + 0.12156262248754501, + 0.03586680814623833, + 0.28840506076812744, + 0.3254171907901764 + ], + "credit_map": [ + 0.009591704234480858, + 0.01084472332149744, + 0.013038937002420425, + 0.009934675879776478, + 0.00920910481363535, + 0.006310689263045788, + 0.008175976574420929, + 0.007040719501674175, + 0.004948272369801998, + 0.003222494386136532, + 0.006827241275459528, + 0.004372435621917248, + 0.003456651931628585, + 0.0037350193597376347, + 0.003567169886082411, + 0.0038183717988431454, + 0.00321621960029006, + 0.0011754895094782114, + 0.005055389832705259, + 0.001968041993677616, + 0.0026079805102199316, + 0.0016585157718509436, + 0.0030774944461882114, + 0.003150918520987034, + 0.0035419848281890154, + 0.0023288559168577194, + 0.0027571481186896563, + 0.0021600788459181786, + 0.002000887878239155, + 0.0024387394078075886, + 0.0022074999287724495, + 0.002232495229691267, + 0.005246436223387718, + 0.0059028975665569305, + 0.009499629959464073, + 0.00407472625374794, + 0.0033594337292015553, + 0.006431706715375185, + 0.005143287125974894, + 0.003526348154991865, + 0.0042138295248150826, + 0.00606371508911252, + 0.008291367441415787, + 0.004858548752963543, + 0.0027008834294974804, + 0.004343263804912567, + 0.002595957601442933, + 0.0028243386186659336, + 0.008854759857058525, + 0.041403938084840775, + 0.03531919792294502, + 0.007418587803840637, + 0.006569772958755493, + 0.004398751072585583, + 0.008483841083943844, + 0.009113765321671963, + 0.007025982718914747, + 0.004212441388517618, + 0.010221663862466812, + 0.004369278438389301, + 0.0019597699865698814, + 0.0029610441997647285, + 0.002802202943712473, + 0.002314424142241478, + 0.06588911265134811, + 0.019070954993367195, + 0.07634564489126205, + 0.02758004702627659, + 0.03007504902780056, + 0.031244348734617233, + 0.015719447284936905, + 0.022480476647615433, + 0.05359206348657608, + 0.10471471399068832, + 0.03117293119430542, + 0.0392686165869236, + 0.026063475757837296, + 0.02415631338953972, + 0.02536248415708542, + 0.02108658477663994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.011081461794674397 + }, + { + "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 tube of lipstick with a deep velvety crimson bullet and a smooth glossy black metal casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lipstick positioned vertically against a dark, polished, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lipstick bullet melt, crumble, and slide aside to form a viscous stream of liquefied crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a droplet fall from the melting lipstick to form a small, perfect circle of shimmering liquid on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed with a handheld camera from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6934114098548889 + ], + "opposite_logprob_tokens": [ + -0.6934114098548889 + ], + "logprob_reward": -0.6934114098548889, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934114098548889, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14719262719154358, + 0.14681097865104675, + 0.07040165364742279, + 0.09549747407436371, + 0.08153431117534637, + 0.11907059699296951, + 0.08052260428667068, + 0.06918332725763321, + 0.023226890712976456, + 0.1665596216917038 + ], + "credit_map": [ + 0.0226978100836277, + 0.021291999146342278, + 0.03106215037405491, + 0.014039594680070877, + 0.014122797176241875, + 0.01669619418680668, + 0.013323554769158363, + 0.013958504423499107, + 0.031037691980600357, + 0.012704833410680294, + 0.025485115125775337, + 0.016895128414034843, + 0.012612918391823769, + 0.018584642559289932, + 0.014421014115214348, + 0.015069614164531231, + 0.018485961481928825, + 0.005571245681494474, + 0.012076730839908123, + 0.005991114769130945, + 0.005591037217527628, + 0.00731549970805645, + 0.005479258485138416, + 0.00989079661667347, + 0.014101959764957428, + 0.010299707762897015, + 0.016460826620459557, + 0.012496184557676315, + 0.008756725117564201, + 0.011025778949260712, + 0.01035825815051794, + 0.011998023837804794, + 0.012777180410921574, + 0.010807790793478489, + 0.014241334050893784, + 0.01097255852073431, + 0.005056705325841904, + 0.010488579981029034, + 0.010717282071709633, + 0.006472864653915167, + 0.008998741395771503, + 0.0289605800062418, + 0.03896914795041084, + 0.011380220763385296, + 0.0065979911014437675, + 0.012727432884275913, + 0.007388869300484657, + 0.0040475958958268166, + 0.0114443963393569, + 0.020441649481654167, + 0.01796218939125538, + 0.008068735711276531, + 0.006014410872012377, + 0.003874492133036256, + 0.0054583619348704815, + 0.007258363533765078, + 0.009361298754811287, + 0.009163081645965576, + 0.013133145868778229, + 0.007767440751194954, + 0.006650504656136036, + 0.007786107249557972, + 0.008275657892227173, + 0.007046080660074949, + 0.002582875080406666, + 0.002556810388341546, + 0.002946810331195593, + 0.004239658825099468, + 0.002114334609359503, + 0.0026304498314857483, + 0.0025997417978942394, + 0.003556207986548543, + 0.019531767815351486, + 0.05462770536541939, + 0.02521951124072075, + 0.01792137324810028, + 0.01246842835098505, + 0.013652653433382511, + 0.012127614580094814, + 0.011010559275746346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.6934114098548889 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_tensor([3167, 3167])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_tensor([3167, 3167])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1e004c7fcee8ab1be8f270abb775cad83e07a2f5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_tensor([3167, 3167])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, tightly framed, surreal macro photograph captures a giant chalk beetle, motionless on a vibrant green, dew-covered fern leaf. Soft, diffused morning light filters through the invisible canopy, revealing the stunning, oil-slick iridescence of its massive black carapace, shimmering with subtle hues of deep green and purple. Its three awe-inspiring horns are rendered with exquisite clarity, their surfaces polished and powerful. The beetle's powerful, segmented legs grasp the fern with microscopic precision, each tiny spike and joint articulated in breathtaking, high-fidelity detail. Spherical dewdrops cling to the fern's delicate fronds, each a perfect, shimmering lens reflecting the tranquil morning scene. The background dissolves into a soft, emerald blur, creating an extremely shallow depth of field, isolating the subject in its miniature world and crafting a serene and powerful atmosphere of natural art.", + "seed": 16252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_3167_seed_16252_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": -0.06276772152632475, + "vlm_logprob": -0.06276772152632475, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.06276772152632475, + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21176418662071228 + ], + "opposite_logprob_tokens": [ + -1.7117642164230347 + ], + "logprob_reward": -0.21176418662071228, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21176418662071228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325482428073883, + 0.07200193405151367, + 0.03589540347456932, + 0.0607389435172081, + 0.050199154764413834, + 0.11877482384443283, + 0.16322775185108185, + 0.03358282521367073, + 0.2836645841598511, + 0.04936635494232178 + ], + "credit_map": [ + 0.021972090005874634, + 0.01997506618499756, + 0.016429075971245766, + 0.019338320940732956, + 0.01602010987699032, + 0.009740124456584454, + 0.010667746886610985, + 0.01840570755302906, + 0.012849710881710052, + 0.005695341620594263, + 0.008766676299273968, + 0.006208114791661501, + 0.005560361780226231, + 0.008166597224771976, + 0.008276143111288548, + 0.01647898368537426, + 0.007400385569781065, + 0.004921623971313238, + 0.0048375437036156654, + 0.0033846490550786257, + 0.003164730966091156, + 0.0024946467019617558, + 0.0034071565605700016, + 0.006284669507294893, + 0.018786972388625145, + 0.0029708927031606436, + 0.004287729039788246, + 0.004067383240908384, + 0.0031758162658661604, + 0.0053752632811665535, + 0.004823198076337576, + 0.017251690849661827, + 0.005629393272101879, + 0.007959366776049137, + 0.006223675794899464, + 0.0066526541486382484, + 0.00546640157699585, + 0.006763537414371967, + 0.006264904513955116, + 0.005239220801740885, + 0.020791027694940567, + 0.01876206137239933, + 0.009822935797274113, + 0.01721327379345894, + 0.016221724450588226, + 0.006312347948551178, + 0.012911698780953884, + 0.016739754006266594, + 0.043664030730724335, + 0.01506663765758276, + 0.00850085262209177, + 0.01437005028128624, + 0.011588037014007568, + 0.012687931768596172, + 0.019124524667859077, + 0.0382256880402565, + 0.004640315659344196, + 0.003697303356602788, + 0.0034180753864347935, + 0.003744696266949177, + 0.003852890571579337, + 0.004171971697360277, + 0.0045348177663981915, + 0.005522754043340683, + 0.10759716480970383, + 0.010699069127440453, + 0.1148654967546463, + 0.012149869464337826, + 0.0036981829907745123, + 0.005058650393038988, + 0.0074121588841080666, + 0.02218397706747055, + 0.010745999403297901, + 0.004429154563695192, + 0.01274518109858036, + 0.003265853738412261, + 0.003117451909929514, + 0.004535364452749491, + 0.004372889641672373, + 0.006154458969831467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.21176418662071228 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006765435915440321 + ], + "opposite_logprob_tokens": [ + -5.006765365600586 + ], + "logprob_reward": -0.006765435915440321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006765435915440321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11225947737693787, + 0.05654136836528778, + 0.09948419779539108, + 0.07149031013250351, + 0.04324285313487053, + 0.041175708174705505, + 0.08182458579540253, + 0.06903093308210373, + 0.36342185735702515, + 0.06152864173054695 + ], + "credit_map": [ + 0.019111819565296173, + 0.012717717327177525, + 0.014227514155209064, + 0.014450492337346077, + 0.01446064654737711, + 0.013543212786316872, + 0.011056276969611645, + 0.01269179955124855, + 0.008351348340511322, + 0.005300928372889757, + 0.005584762431681156, + 0.007526956498622894, + 0.0070185791701078415, + 0.008969301357865334, + 0.005081984680145979, + 0.008707508444786072, + 0.01679270900785923, + 0.009442764334380627, + 0.011380818672478199, + 0.013929584063589573, + 0.014493056572973728, + 0.01171054132282734, + 0.00973964761942625, + 0.01199508085846901, + 0.01809542626142502, + 0.005033331457525492, + 0.005161081440746784, + 0.0037694782949984074, + 0.005945684853941202, + 0.007016997318714857, + 0.007144575472921133, + 0.01932373270392418, + 0.006423704791814089, + 0.00441794702783227, + 0.0050141774117946625, + 0.00484856404364109, + 0.004469463601708412, + 0.0041488646529614925, + 0.006520349066704512, + 0.007399783004075289, + 0.005573288071900606, + 0.006422452628612518, + 0.003788638859987259, + 0.005726953502744436, + 0.006117918528616428, + 0.0021172771230340004, + 0.006426267791539431, + 0.00500291446223855, + 0.012523798272013664, + 0.009128482080996037, + 0.007899407297372818, + 0.009074573405086994, + 0.009463376365602016, + 0.010302496142685413, + 0.01315562054514885, + 0.01027682889252901, + 0.008570854552090168, + 0.005322462413460016, + 0.009147978387773037, + 0.005347858183085918, + 0.009333309717476368, + 0.00686861714348197, + 0.00995955616235733, + 0.014480296522378922, + 0.1098208874464035, + 0.012837878428399563, + 0.14430001378059387, + 0.018514584749937057, + 0.007258850149810314, + 0.012333433143794537, + 0.012221354059875011, + 0.04613484814763069, + 0.015871677547693253, + 0.0039779203943908215, + 0.019277596846222878, + 0.005105920135974884, + 0.0028955505695194006, + 0.004134685732424259, + 0.004891990218311548, + 0.005373301915824413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.006765435915440321 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041004871018230915 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.0041004871018230915, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041004871018230915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13877815008163452, + 0.06974947452545166, + 0.039966098964214325, + 0.039168789982795715, + 0.05547064170241356, + 0.03972295671701431, + 0.2512587606906891, + 0.045100100338459015, + 0.27497056126594543, + 0.04581451416015625 + ], + "credit_map": [ + 0.020195581018924713, + 0.014359858818352222, + 0.013599184341728687, + 0.022872447967529297, + 0.023294758051633835, + 0.014582431875169277, + 0.013811670243740082, + 0.016062220558524132, + 0.009292266331613064, + 0.006923504639416933, + 0.006894960533827543, + 0.01088634692132473, + 0.010988126508891582, + 0.008150189183652401, + 0.007148416247218847, + 0.009465659037232399, + 0.007777581922709942, + 0.004566325340420008, + 0.00462944433093071, + 0.006832081358879805, + 0.0034187843557447195, + 0.004601019900292158, + 0.0034675286151468754, + 0.004673334304243326, + 0.007291007786989212, + 0.0032658614218235016, + 0.003847162239253521, + 0.005297503434121609, + 0.005124344956129789, + 0.003535959403961897, + 0.005338015966117382, + 0.005468936171382666, + 0.008128619752824306, + 0.007515641860663891, + 0.006505188997834921, + 0.007286832667887211, + 0.005398913286626339, + 0.00671053770929575, + 0.004360470920801163, + 0.009564436040818691, + 0.00528583163395524, + 0.0071917870081961155, + 0.004020333290100098, + 0.006361231207847595, + 0.005579735618084669, + 0.002640068531036377, + 0.004387706983834505, + 0.004256261978298426, + 0.04914300516247749, + 0.02134396880865097, + 0.014429434202611446, + 0.03865155950188637, + 0.024558980017900467, + 0.021584900096058846, + 0.02988104335963726, + 0.051665887236595154, + 0.005394205916672945, + 0.004549541510641575, + 0.004777909722179174, + 0.00392297375947237, + 0.007965977303683758, + 0.0063724578358232975, + 0.007085962686687708, + 0.0050310720689594746, + 0.0883331447839737, + 0.010468912310898304, + 0.10901521146297455, + 0.013250540010631084, + 0.005070369690656662, + 0.006363645661622286, + 0.0092572420835495, + 0.033211495727300644, + 0.006618117913603783, + 0.006120918784290552, + 0.007904930040240288, + 0.0060392809100449085, + 0.004866667091846466, + 0.003465682500973344, + 0.006005400791764259, + 0.004793513100594282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0041004871018230915 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901328593492508 + ], + "opposite_logprob_tokens": [ + -2.4640133380889893 + ], + "logprob_reward": -0.08901328593492508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901328593492508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10109110176563263, + 0.08443435281515121, + 0.061724916100502014, + 0.05701908469200134, + 0.07711721211671829, + 0.06684602051973343, + 0.06685721129179001, + 0.11626026034355164, + 0.3004269003868103, + 0.06822296977043152 + ], + "credit_map": [ + 0.016795022413134575, + 0.012257938273251057, + 0.015037248842418194, + 0.01565776765346527, + 0.014848022721707821, + 0.008036862127482891, + 0.006956868339329958, + 0.011501376517117023, + 0.013288804329931736, + 0.009980542585253716, + 0.011693956330418587, + 0.01363462395966053, + 0.010977459140121937, + 0.008762815967202187, + 0.007474118378013372, + 0.008622036315500736, + 0.00826798565685749, + 0.011800610460340977, + 0.006569110322743654, + 0.010584172792732716, + 0.00831483956426382, + 0.005081990733742714, + 0.005174768157303333, + 0.0059314328245818615, + 0.020271150395274162, + 0.0029021890368312597, + 0.006428826600313187, + 0.002708942163735628, + 0.002824928844347596, + 0.003917814698070288, + 0.0037549613043665886, + 0.014210275374352932, + 0.007123463321477175, + 0.013123546727001667, + 0.008822312578558922, + 0.012118425220251083, + 0.010789385065436363, + 0.006856150459498167, + 0.008255685679614544, + 0.010028245858848095, + 0.00923091359436512, + 0.008800390176475048, + 0.006835341919213533, + 0.01123131811618805, + 0.008143986575305462, + 0.004436211194843054, + 0.009518281556665897, + 0.008649574592709541, + 0.0073213130235672, + 0.007900363765656948, + 0.005340625066310167, + 0.0092701381072402, + 0.009325360879302025, + 0.006413953844457865, + 0.012402351014316082, + 0.008883105590939522, + 0.014327647164463997, + 0.015154398046433926, + 0.012484137900173664, + 0.013862655498087406, + 0.020764315500855446, + 0.01251815352588892, + 0.014244629070162773, + 0.01290432084351778, + 0.08717021346092224, + 0.017698343843221664, + 0.10870454460382462, + 0.014700736850500107, + 0.010674075223505497, + 0.011274587363004684, + 0.010341458953917027, + 0.03986295312643051, + 0.008689765818417072, + 0.008244565688073635, + 0.007013075985014439, + 0.009882125072181225, + 0.0072473869659006596, + 0.009792102500796318, + 0.009468864649534225, + 0.007885083556175232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.08901328593492508 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195212058722973 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195212058722973, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195212058722973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719319075345993, + 0.05229458212852478, + 0.044885553419589996, + 0.07720081508159637, + 0.03495658561587334, + 0.018200919032096863, + 0.06185506284236908, + 0.11649929732084274, + 0.3861246109008789, + 0.0360507033765316 + ], + "credit_map": [ + 0.024556206539273262, + 0.0171066801995039, + 0.016860684379935265, + 0.021069442853331566, + 0.027941621840000153, + 0.020795127376914024, + 0.017017647624015808, + 0.026584487408399582, + 0.008989451453089714, + 0.004334552213549614, + 0.006894342601299286, + 0.005276505369693041, + 0.00491303950548172, + 0.0064527601934969425, + 0.004441891796886921, + 0.010992041788995266, + 0.007977179251611233, + 0.0036061431746929884, + 0.005490429699420929, + 0.004401210695505142, + 0.004289261996746063, + 0.007641550153493881, + 0.004766702651977539, + 0.006713074631989002, + 0.011605209670960903, + 0.006991338916122913, + 0.00956945214420557, + 0.007033111527562141, + 0.00958920456469059, + 0.010725612752139568, + 0.011223625391721725, + 0.010463257320225239, + 0.005357641726732254, + 0.003955731634050608, + 0.004747265018522739, + 0.003037894843146205, + 0.003435723716393113, + 0.004398086108267307, + 0.004929625894874334, + 0.0050946190021932125, + 0.0028299158439040184, + 0.0016791853122413158, + 0.002942768856883049, + 0.0017613042145967484, + 0.0020687070209532976, + 0.0022184227127581835, + 0.0018776500364765525, + 0.002822963520884514, + 0.015423440374433994, + 0.004232689738273621, + 0.0043912664987146854, + 0.003381829708814621, + 0.0038157766684889793, + 0.008584496565163136, + 0.006382775958627462, + 0.015642788261175156, + 0.014782960526645184, + 0.011670736595988274, + 0.021957240998744965, + 0.01198381558060646, + 0.01269806083291769, + 0.015256084501743317, + 0.015163279138505459, + 0.012987109832465649, + 0.1357332468032837, + 0.01499956101179123, + 0.15597820281982422, + 0.017604639753699303, + 0.006453884765505791, + 0.008392183110117912, + 0.012137466110289097, + 0.03482537716627121, + 0.0046590170823037624, + 0.004445680417120457, + 0.0037085076328366995, + 0.004367448389530182, + 0.002898711711168289, + 0.00591106666252017, + 0.004376744851469994, + 0.005683530122041702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.002195212058722973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, tightly framed, surreal macro photograph captures a giant chalk beetle, motionless on a vibrant green, dew-covered fern leaf. Soft, diffused morning light filters through the invisible canopy, revealing the stunning, oil-slick iridescence of its massive black carapace, shimmering with subtle hues of deep green and purple. Its three awe-inspiring horns are rendered with exquisite clarity, their surfaces polished and powerful. The beetle's powerful, segmented legs grasp the fern with microscopic precision, each tiny spike and joint articulated in breathtaking, high-fidelity detail. Spherical dewdrops cling to the fern's delicate fronds, each a perfect, shimmering lens reflecting the tranquil morning scene. The background dissolves into a soft, emerald blur, creating an extremely shallow depth of field, isolating the subject in its miniature world and crafting a serene and powerful atmosphere of natural art.", + "seed": 16253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_3167_seed_16253_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": -0.0480688554700464, + "vlm_logprob": -0.0480688554700464, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.0480688554700464, + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14715978503227234 + ], + "opposite_logprob_tokens": [ + -2.0221598148345947 + ], + "logprob_reward": -0.14715978503227234, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14715978503227234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24431557953357697, + 0.09445265680551529, + 0.05323624983429909, + 0.12553803622722626, + 0.047238245606422424, + 0.03338334336876869, + 0.07731280475854874, + 0.2150116115808487, + 0.07939647883176804, + 0.030114958062767982 + ], + "credit_map": [ + 0.031535059213638306, + 0.03339037671685219, + 0.0406864732503891, + 0.034924980252981186, + 0.03202231973409653, + 0.024499984458088875, + 0.024679770693182945, + 0.022576626390218735, + 0.008290720172226429, + 0.01281476765871048, + 0.014182918705046177, + 0.018809234723448753, + 0.015289646573364735, + 0.007101120427250862, + 0.009023810736835003, + 0.008940445259213448, + 0.007637042086571455, + 0.005176630802452564, + 0.008478205651044846, + 0.008160618133842945, + 0.006117744836956263, + 0.0057206833735108376, + 0.006965452805161476, + 0.004979879129678011, + 0.011794847436249256, + 0.013290519826114178, + 0.016774168238043785, + 0.015962202101945877, + 0.02094116434454918, + 0.008977008052170277, + 0.024859631434082985, + 0.012938513420522213, + 0.005443489644676447, + 0.005076339468359947, + 0.007166517898440361, + 0.005411824677139521, + 0.006290525663644075, + 0.005548093467950821, + 0.006168241146951914, + 0.006133217364549637, + 0.004468542989343405, + 0.0029597494285553694, + 0.0047405678778886795, + 0.0034082301426678896, + 0.003021880751475692, + 0.005833829287439585, + 0.004034040495753288, + 0.004916505888104439, + 0.008420192636549473, + 0.004814467392861843, + 0.010844754986464977, + 0.008836191147565842, + 0.006209483835846186, + 0.014355678111314774, + 0.013827110640704632, + 0.010004929266870022, + 0.035942163318395615, + 0.02020367421209812, + 0.03489173576235771, + 0.029489073902368546, + 0.022377248853445053, + 0.02652617357671261, + 0.016879605129361153, + 0.02870194800198078, + 0.010736026801168919, + 0.0063239349983632565, + 0.011968794278800488, + 0.007744870614260435, + 0.009707871824502945, + 0.010272924788296223, + 0.01368706300854683, + 0.008954985998570919, + 0.0060755363665521145, + 0.0027534605469554663, + 0.004028202034533024, + 0.003736893180757761, + 0.0019219157984480262, + 0.004009215161204338, + 0.003671131329610944, + 0.003918607719242573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.14715978503227234 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647995837032795 + ], + "opposite_logprob_tokens": [ + -4.88264799118042 + ], + "logprob_reward": -0.007647995837032795, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647995837032795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11303795129060745, + 0.11285576969385147, + 0.07224704325199127, + 0.04411779344081879, + 0.07345189899206161, + 0.06331239640712738, + 0.09728430956602097, + 0.08509089797735214, + 0.20840167999267578, + 0.13020026683807373 + ], + "credit_map": [ + 0.016254296526312828, + 0.01653408445417881, + 0.01988758146762848, + 0.01569763571023941, + 0.010939973406493664, + 0.009402486495673656, + 0.011566282249987125, + 0.012755615636706352, + 0.013716001994907856, + 0.01141207292675972, + 0.02147100865840912, + 0.020623788237571716, + 0.010290847159922123, + 0.010214569047093391, + 0.012822875753045082, + 0.012304609641432762, + 0.007895690388977528, + 0.010425001382827759, + 0.009625675156712532, + 0.011847153306007385, + 0.010556950233876705, + 0.006610750686377287, + 0.0075584533624351025, + 0.007727370131760836, + 0.006209352519363165, + 0.005205079447478056, + 0.006827539764344692, + 0.006815833505243063, + 0.004715498071163893, + 0.004563318565487862, + 0.004779499024152756, + 0.0050016725435853004, + 0.01034520287066698, + 0.007979005575180054, + 0.011885212734341621, + 0.009147039614617825, + 0.009163576178252697, + 0.006618062034249306, + 0.009249070659279823, + 0.009064733982086182, + 0.008470697328448296, + 0.0071822768077254295, + 0.008773240260779858, + 0.008784536272287369, + 0.0052518113516271114, + 0.009371291846036911, + 0.0069938162341713905, + 0.00848473235964775, + 0.01484186016023159, + 0.01062180008739233, + 0.014363457448780537, + 0.012863107956945896, + 0.008559362962841988, + 0.011173435486853123, + 0.014963139779865742, + 0.009898136369884014, + 0.01610146276652813, + 0.006698375102132559, + 0.010963479988276958, + 0.013756300322711468, + 0.010321867652237415, + 0.00980502087622881, + 0.007772552780807018, + 0.009671831503510475, + 0.030106201767921448, + 0.023389644920825958, + 0.03333606570959091, + 0.018862714990973473, + 0.018707752227783203, + 0.026896808296442032, + 0.020980408415198326, + 0.03612208738923073, + 0.018515964969992638, + 0.010903853923082352, + 0.018709272146224976, + 0.01321145799010992, + 0.011692785657942295, + 0.027442291378974915, + 0.012619505636394024, + 0.01710512861609459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.007647995837032795 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097281955182552 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097281955182552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097281955182552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1689567118883133, + 0.13773976266384125, + 0.11170436441898346, + 0.0582626536488533, + 0.0405227467417717, + 0.03949744999408722, + 0.09318070858716965, + 0.07275737822055817, + 0.20524489879608154, + 0.07213331758975983 + ], + "credit_map": [ + 0.02568494714796543, + 0.029554899781942368, + 0.026395371183753014, + 0.01948273926973343, + 0.01630936749279499, + 0.015725867822766304, + 0.018089059740304947, + 0.017714451998472214, + 0.015077808871865273, + 0.021522395312786102, + 0.024549033492803574, + 0.01804894581437111, + 0.01670190691947937, + 0.00922281201928854, + 0.01639457978308201, + 0.016222290694713593, + 0.0098954476416111, + 0.019890232011675835, + 0.013326456770300865, + 0.025503545999526978, + 0.012302303686738014, + 0.007441475987434387, + 0.011052149347960949, + 0.012292761355638504, + 0.0066873678006231785, + 0.009264742024242878, + 0.007403080817312002, + 0.007268836256116629, + 0.008271762169897556, + 0.00663736741989851, + 0.007000564131885767, + 0.005728934891521931, + 0.005697100423276424, + 0.005285289604216814, + 0.005246943794190884, + 0.005137267056852579, + 0.0044756606221199036, + 0.0038626608438789845, + 0.006279313005506992, + 0.004538513254374266, + 0.005752891767770052, + 0.004135949537158012, + 0.005372301209717989, + 0.0050234198570251465, + 0.0035099599044770002, + 0.006013114936649799, + 0.004275517538189888, + 0.005414295010268688, + 0.00983375683426857, + 0.00817197747528553, + 0.012700035236775875, + 0.011358354240655899, + 0.008589302189648151, + 0.020043475553393364, + 0.011312785558402538, + 0.011171023361384869, + 0.008562606759369373, + 0.006895874161273241, + 0.00906949769705534, + 0.014873025007545948, + 0.006608948577195406, + 0.010288518853485584, + 0.00793435238301754, + 0.008524550125002861, + 0.020535094663500786, + 0.019730379804968834, + 0.029624171555042267, + 0.017340151593089104, + 0.025478806346654892, + 0.03635594621300697, + 0.028867485001683235, + 0.0273128729313612, + 0.008832499384880066, + 0.005585680715739727, + 0.009666337631642818, + 0.010546456091105938, + 0.006507298443466425, + 0.012122814543545246, + 0.01018207985907793, + 0.008690147660672665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.004097281955182552 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895452529191971 + ], + "opposite_logprob_tokens": [ + -2.5789544582366943 + ], + "logprob_reward": -0.07895452529191971, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895452529191971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13978660106658936, + 0.07056319713592529, + 0.04053905978798866, + 0.2232106477022171, + 0.09066252410411835, + 0.03706718981266022, + 0.08743058890104294, + 0.12480363994836807, + 0.12823159992694855, + 0.05770498886704445 + ], + "credit_map": [ + 0.020546795800328255, + 0.01684682071208954, + 0.024040140211582184, + 0.01746133156120777, + 0.01352416630834341, + 0.013883513398468494, + 0.015346302650868893, + 0.018137522041797638, + 0.00719377351924777, + 0.008188873529434204, + 0.012007453478872776, + 0.012512759305536747, + 0.01009221002459526, + 0.004853701684623957, + 0.008757099509239197, + 0.006957318168133497, + 0.004781341180205345, + 0.004832280334085226, + 0.0055165658704936504, + 0.006361057981848717, + 0.0041455780155956745, + 0.0063665262423455715, + 0.004194715991616249, + 0.004340992774814367, + 0.031523481011390686, + 0.018367953598499298, + 0.035024382174015045, + 0.024269597604870796, + 0.03457684814929962, + 0.020363541319966316, + 0.033502642065286636, + 0.02558218128979206, + 0.017586467787623405, + 0.008437879383563995, + 0.012436842545866966, + 0.007521297316998243, + 0.012999075464904308, + 0.008325747214257717, + 0.012841848656535149, + 0.01051335595548153, + 0.005123603623360395, + 0.003109327983111143, + 0.006306011229753494, + 0.003597702132537961, + 0.005231703631579876, + 0.004491832107305527, + 0.004175415728241205, + 0.005031589884310961, + 0.007702535949647427, + 0.009677302092313766, + 0.012166001833975315, + 0.010376804508268833, + 0.007846982218325138, + 0.01859973929822445, + 0.010498898103833199, + 0.010562322102487087, + 0.013196201995015144, + 0.009551196359097958, + 0.020583879202604294, + 0.018134700134396553, + 0.012971744872629642, + 0.016260512173175812, + 0.010752289555966854, + 0.02335311472415924, + 0.013608803041279316, + 0.014751007780432701, + 0.014741283841431141, + 0.01187728438526392, + 0.019158633425831795, + 0.023312801495194435, + 0.016027698293328285, + 0.014754069037735462, + 0.010104818269610405, + 0.005298267584294081, + 0.0073342351242899895, + 0.007853555493056774, + 0.004694770555943251, + 0.008373502641916275, + 0.007195476442575455, + 0.00685036089271307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.07895452529191971 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024846892338246107 + ], + "opposite_logprob_tokens": [ + -6.0024847984313965 + ], + "logprob_reward": -0.0024846892338246107, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024846892338246107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08558627963066101, + 0.07436003535985947, + 0.046466682106256485, + 0.1655118316411972, + 0.08115357160568237, + 0.061131592839956284, + 0.14007055759429932, + 0.09528569877147675, + 0.11709288507699966, + 0.13334089517593384 + ], + "credit_map": [ + 0.011912243440747261, + 0.008740912191569805, + 0.013460749760270119, + 0.011937650851905346, + 0.008917619474232197, + 0.011296859942376614, + 0.009211740456521511, + 0.01010849792510271, + 0.009065431542694569, + 0.007581210229545832, + 0.013412046246230602, + 0.009095038287341595, + 0.006562840659171343, + 0.007582636084407568, + 0.012218411080539227, + 0.008842410519719124, + 0.005802912637591362, + 0.004971855785697699, + 0.0052131167612969875, + 0.0058443984016776085, + 0.0067160846665501595, + 0.006429597269743681, + 0.006418148521333933, + 0.005070561543107033, + 0.018536539748311043, + 0.020807119086384773, + 0.022420531138777733, + 0.02212425135076046, + 0.023955613374710083, + 0.016874969005584717, + 0.023432834073901176, + 0.017359955236315727, + 0.015343629755079746, + 0.005701512098312378, + 0.01326408889144659, + 0.006771148648113012, + 0.00975489430129528, + 0.00951344147324562, + 0.012045812793076038, + 0.008759040385484695, + 0.00750203151255846, + 0.006290072575211525, + 0.010548427700996399, + 0.0065923635847866535, + 0.006153172813355923, + 0.009419857524335384, + 0.006383445113897324, + 0.00824221596121788, + 0.015051344409584999, + 0.0074624670669436455, + 0.0189807191491127, + 0.01565169170498848, + 0.01049394067376852, + 0.027770185843110085, + 0.025542374700307846, + 0.019117826595902443, + 0.012195915915071964, + 0.00993520300835371, + 0.011335151270031929, + 0.017399340867996216, + 0.01112220250070095, + 0.013033543713390827, + 0.010199112817645073, + 0.010065220296382904, + 0.01484336331486702, + 0.010089388117194176, + 0.01899985782802105, + 0.011196577921509743, + 0.010377136059105396, + 0.016530700027942657, + 0.014495203271508217, + 0.020560644567012787, + 0.01938037946820259, + 0.009716992266476154, + 0.019947903230786324, + 0.01766069233417511, + 0.01001778244972229, + 0.025106659159064293, + 0.01528116688132286, + 0.016229311004281044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.0024846892338246107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, tightly framed, surreal macro photograph captures a giant chalk beetle, motionless on a vibrant green, dew-covered fern leaf. Soft, diffused morning light filters through the invisible canopy, revealing the stunning, oil-slick iridescence of its massive black carapace, shimmering with subtle hues of deep green and purple. Its three awe-inspiring horns are rendered with exquisite clarity, their surfaces polished and powerful. The beetle's powerful, segmented legs grasp the fern with microscopic precision, each tiny spike and joint articulated in breathtaking, high-fidelity detail. Spherical dewdrops cling to the fern's delicate fronds, each a perfect, shimmering lens reflecting the tranquil morning scene. The background dissolves into a soft, emerald blur, creating an extremely shallow depth of field, isolating the subject in its miniature world and crafting a serene and powerful atmosphere of natural art.", + "seed": 16254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_3167_seed_16254_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": -0.04918570942245424, + "vlm_logprob": -0.04918570942245424, + "vlm_yes_no_margin": 4.349999976158142, + "vlm_token_credit": -0.04918570942245424, + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13428623974323273 + ], + "opposite_logprob_tokens": [ + -2.1342861652374268 + ], + "logprob_reward": -0.13428623974323273, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13428623974323273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2016363888978958, + 0.07329849898815155, + 0.07745884358882904, + 0.04020991548895836, + 0.033360157161951065, + 0.0719858705997467, + 0.08722412586212158, + 0.054897211492061615, + 0.1650470644235611, + 0.1948818415403366 + ], + "credit_map": [ + 0.025201017037034035, + 0.01855970360338688, + 0.021551057696342468, + 0.01411427278071642, + 0.025697221979498863, + 0.03307439759373665, + 0.019262205809354782, + 0.044176530092954636, + 0.014538885094225407, + 0.0068311551585793495, + 0.01318024005740881, + 0.00857328251004219, + 0.005653941538184881, + 0.0064104702323675156, + 0.00945474486798048, + 0.008655786514282227, + 0.01315508596599102, + 0.010039507411420345, + 0.010755261406302452, + 0.0071776448749005795, + 0.011839508078992367, + 0.006575970910489559, + 0.008006228134036064, + 0.009909640066325665, + 0.006239392329007387, + 0.0038302617613226175, + 0.006839266512542963, + 0.004237475339323282, + 0.004916332196444273, + 0.0026290544774383307, + 0.004709975328296423, + 0.006808159407228231, + 0.006943911779671907, + 0.003113304963335395, + 0.0058824848383665085, + 0.0035868685226887465, + 0.003378977533429861, + 0.0025989862624555826, + 0.0027760867960751057, + 0.005079539958387613, + 0.008271737955510616, + 0.016197865828871727, + 0.009062830358743668, + 0.008187266066670418, + 0.010718048550188541, + 0.004801883827894926, + 0.00625916663557291, + 0.008487085811793804, + 0.018310831859707832, + 0.007835021242499352, + 0.022725211456418037, + 0.00843650009483099, + 0.005826377309858799, + 0.005007152445614338, + 0.008899075910449028, + 0.010183961130678654, + 0.005292159970849752, + 0.009170926176011562, + 0.008203276433050632, + 0.005202363710850477, + 0.008981880731880665, + 0.004428265616297722, + 0.004727435763925314, + 0.008890909142792225, + 0.02908628061413765, + 0.010221763513982296, + 0.04931065812706947, + 0.012795032933354378, + 0.0142832575365901, + 0.015668218955397606, + 0.0090176435187459, + 0.024664226919412613, + 0.023804739117622375, + 0.02731063775718212, + 0.03680172190070152, + 0.023725686594843864, + 0.025256343185901642, + 0.023995783179998398, + 0.018544897437095642, + 0.015442046336829662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.13428623974323273 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00765107199549675 + ], + "opposite_logprob_tokens": [ + -4.882650852203369 + ], + "logprob_reward": -0.00765107199549675, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00765107199549675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20505499839782715, + 0.09582715481519699, + 0.10238655656576157, + 0.03703933209180832, + 0.05840107053518295, + 0.04372049495577812, + 0.07892056554555893, + 0.06338029354810715, + 0.14030534029006958, + 0.17496420443058014 + ], + "credit_map": [ + 0.04414449259638786, + 0.018176017329096794, + 0.035546544939279556, + 0.023030536249279976, + 0.019566675648093224, + 0.02196526899933815, + 0.021571125835180283, + 0.021054353564977646, + 0.021756064146757126, + 0.011975592002272606, + 0.01156863197684288, + 0.0073982225731015205, + 0.010175427421927452, + 0.007977873086929321, + 0.008978667668998241, + 0.015996679663658142, + 0.016907820478081703, + 0.007378028240054846, + 0.018232837319374084, + 0.011503470130264759, + 0.012013547122478485, + 0.005893426015973091, + 0.01333415787667036, + 0.017123263329267502, + 0.009176862426102161, + 0.0034183936659246683, + 0.0051643517799675465, + 0.003259809920564294, + 0.002637996803969145, + 0.0023151051718741655, + 0.0034295946825295687, + 0.007637217175215483, + 0.010101302526891232, + 0.006668953225016594, + 0.009608738124370575, + 0.0086003253236413, + 0.006442859303206205, + 0.00384542322717607, + 0.005159628111869097, + 0.007973841391503811, + 0.005165225360542536, + 0.004560539964586496, + 0.006720326375216246, + 0.00835732277482748, + 0.0049771093763411045, + 0.002570169745013118, + 0.006161005701869726, + 0.005208794958889484, + 0.016929415985941887, + 0.00782093033194542, + 0.011356745846569538, + 0.007298202253878117, + 0.00898923072963953, + 0.008911048993468285, + 0.007776439655572176, + 0.009838548488914967, + 0.008601519279181957, + 0.0071230255998671055, + 0.013344619423151016, + 0.0057443310506641865, + 0.009519106708467007, + 0.00557154044508934, + 0.004549285396933556, + 0.008926868438720703, + 0.026804119348526, + 0.007058556657284498, + 0.0467509888112545, + 0.008904545567929745, + 0.010020969435572624, + 0.010926852934062481, + 0.005943412892520428, + 0.023895887657999992, + 0.019025025889277458, + 0.018854795023798943, + 0.024392519146203995, + 0.018108531832695007, + 0.026145465672016144, + 0.023740340024232864, + 0.01621018722653389, + 0.028487352654337883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00765107199549675 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200802555307746 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200802555307746, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200802555307746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16248588263988495, + 0.05391528829932213, + 0.028768880292773247, + 0.06766040623188019, + 0.038034260272979736, + 0.051939141005277634, + 0.07071199268102646, + 0.061673298478126526, + 0.32324936985969543, + 0.14156146347522736 + ], + "credit_map": [ + 0.020815270021557808, + 0.015090018510818481, + 0.020812781527638435, + 0.012201360426843166, + 0.022883528843522072, + 0.02966334857046604, + 0.01382952369749546, + 0.02719005011022091, + 0.005796641111373901, + 0.0057543800212442875, + 0.006762943230569363, + 0.00981835462152958, + 0.007546004839241505, + 0.005178322549909353, + 0.006530465092509985, + 0.006528180092573166, + 0.004979393910616636, + 0.003308527637273073, + 0.00601885374635458, + 0.0027866396121680737, + 0.0029163614381104708, + 0.002376099117100239, + 0.002746191807091236, + 0.003636815818026662, + 0.02215755544602871, + 0.004524846561253071, + 0.009894582442939281, + 0.003718155901879072, + 0.0039482517167925835, + 0.00462033087387681, + 0.005069372244179249, + 0.013727315701544285, + 0.006631050258874893, + 0.004793947096914053, + 0.006031177472323179, + 0.004814598709344864, + 0.005449492018669844, + 0.0027623556088656187, + 0.002824092283844948, + 0.004727544728666544, + 0.006135546136647463, + 0.0065863835625350475, + 0.00783591065555811, + 0.010780504904687405, + 0.006234488449990749, + 0.00308225373737514, + 0.005599868483841419, + 0.0056841811165213585, + 0.01765456795692444, + 0.005899115465581417, + 0.011488729156553745, + 0.006424689199775457, + 0.008990745060145855, + 0.006127391010522842, + 0.004998419899493456, + 0.009128335863351822, + 0.007832635194063187, + 0.006660772021859884, + 0.012541412375867367, + 0.006463261786848307, + 0.00959805678576231, + 0.005153834819793701, + 0.004152549896389246, + 0.009270777925848961, + 0.05564625933766365, + 0.016763323917984962, + 0.10234237462282181, + 0.02505543641746044, + 0.02963242679834366, + 0.0233700480312109, + 0.015593321062624454, + 0.05484617501497269, + 0.01849096640944481, + 0.014395848847925663, + 0.023963283747434616, + 0.013049831613898277, + 0.020664425566792488, + 0.020368268713355064, + 0.011736692860722542, + 0.018892139196395874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.002200802555307746 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027904808521271 + ], + "opposite_logprob_tokens": [ + -2.3502790927886963 + ], + "logprob_reward": -0.10027904808521271, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027904808521271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16494259238243103, + 0.089627206325531, + 0.06023992598056793, + 0.06961729377508163, + 0.04021265730261803, + 0.06084566190838814, + 0.14053764939308167, + 0.11323411017656326, + 0.11762391775846481, + 0.14311900734901428 + ], + "credit_map": [ + 0.03726271167397499, + 0.014736032113432884, + 0.03003687411546707, + 0.014237025752663612, + 0.016643034294247627, + 0.01905815862119198, + 0.015618279576301575, + 0.01735047809779644, + 0.011677192524075508, + 0.014838621951639652, + 0.015038082376122475, + 0.012102076783776283, + 0.010554018430411816, + 0.00605303468182683, + 0.010475694201886654, + 0.00888848677277565, + 0.01028712373226881, + 0.007677403744310141, + 0.006748131942003965, + 0.006834312807768583, + 0.009760025888681412, + 0.004115632269531488, + 0.0066617196425795555, + 0.008155576884746552, + 0.025841856375336647, + 0.003895650152117014, + 0.009339956566691399, + 0.003158370265737176, + 0.0035636266693472862, + 0.004178206901997328, + 0.003965815529227257, + 0.015673808753490448, + 0.005858574528247118, + 0.005021974910050631, + 0.007996619679033756, + 0.005134559702128172, + 0.00425729900598526, + 0.0022056067828089, + 0.0032635838724672794, + 0.006474437657743692, + 0.007559424266219139, + 0.006747249513864517, + 0.008562379516661167, + 0.009621274657547474, + 0.00812829565256834, + 0.003976902924478054, + 0.007771052420139313, + 0.008479085750877857, + 0.03272498399019241, + 0.012244698591530323, + 0.02350010722875595, + 0.014228248968720436, + 0.018311593681573868, + 0.010864668525755405, + 0.010834568180143833, + 0.017828771844506264, + 0.0130734508857131, + 0.01459089107811451, + 0.0219444427639246, + 0.010479222051799297, + 0.017967507243156433, + 0.009286931715905666, + 0.00898053776472807, + 0.01691112481057644, + 0.021599343046545982, + 0.005293841939419508, + 0.0393054261803627, + 0.00819725263863802, + 0.008625632151961327, + 0.008850581012666225, + 0.005407373886555433, + 0.020344475284218788, + 0.017548277974128723, + 0.015157666988670826, + 0.023076046258211136, + 0.013688331469893456, + 0.02129226364195347, + 0.021049190312623978, + 0.011992977000772953, + 0.019314270466566086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.10027904808521271 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015113847330212593 + ], + "opposite_logprob_tokens": [ + -6.501511573791504 + ], + "logprob_reward": -0.0015113847330212593, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015113847330212593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1067509651184082, + 0.13134269416332245, + 0.09937973320484161, + 0.21873600780963898, + 0.03653654083609581, + 0.046699702739715576, + 0.03990111127495766, + 0.06301167607307434, + 0.04019094258546829, + 0.21745063364505768 + ], + "credit_map": [ + 0.021450882777571678, + 0.010160798206925392, + 0.012024394236505032, + 0.011848405003547668, + 0.013577503152191639, + 0.015572309494018555, + 0.010926001705229282, + 0.011190670542418957, + 0.031017044559121132, + 0.014602912589907646, + 0.014960724860429764, + 0.011105245910584927, + 0.01344750914722681, + 0.010875385254621506, + 0.013239203952252865, + 0.022094685584306717, + 0.014907745644450188, + 0.009524095803499222, + 0.015488874167203903, + 0.009049998596310616, + 0.015448236837983131, + 0.0054122996516525745, + 0.013782679103314877, + 0.015765802934765816, + 0.0701434463262558, + 0.013075168244540691, + 0.03217532858252525, + 0.012594189494848251, + 0.011244445107877254, + 0.01648172177374363, + 0.01600586622953415, + 0.04701584205031395, + 0.006232640706002712, + 0.0034174646716564894, + 0.005617526825517416, + 0.0053886910900473595, + 0.0038873718585819006, + 0.003835115348920226, + 0.003060947870835662, + 0.005096783861517906, + 0.0058353496715426445, + 0.004016533959656954, + 0.009619886055588722, + 0.007724042981863022, + 0.00394166586920619, + 0.003510315902531147, + 0.004481885582208633, + 0.0075700245797634125, + 0.00982736423611641, + 0.0040674940682947636, + 0.005667780060321093, + 0.0035971018951386213, + 0.004972285591065884, + 0.0036246676463633776, + 0.0029367785900831223, + 0.005207641515880823, + 0.01080865878611803, + 0.0068590110167860985, + 0.013049657456576824, + 0.0051406570710241795, + 0.006635131314396858, + 0.0068215979263186455, + 0.005339787341654301, + 0.008357174694538116, + 0.006752170622348785, + 0.0022664640564471483, + 0.011419584043323994, + 0.003195363562554121, + 0.0037736769299954176, + 0.0037318149115890265, + 0.0021490922663360834, + 0.006902776658535004, + 0.02760811522603035, + 0.022531714290380478, + 0.03478376194834709, + 0.020107656717300415, + 0.033136434853076935, + 0.030347341671586037, + 0.017998773604631424, + 0.03093681111931801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.0015113847330212593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, tightly framed, surreal macro photograph captures a giant chalk beetle, motionless on a vibrant green, dew-covered fern leaf. Soft, diffused morning light filters through the invisible canopy, revealing the stunning, oil-slick iridescence of its massive black carapace, shimmering with subtle hues of deep green and purple. Its three awe-inspiring horns are rendered with exquisite clarity, their surfaces polished and powerful. The beetle's powerful, segmented legs grasp the fern with microscopic precision, each tiny spike and joint articulated in breathtaking, high-fidelity detail. Spherical dewdrops cling to the fern's delicate fronds, each a perfect, shimmering lens reflecting the tranquil morning scene. The background dissolves into a soft, emerald blur, creating an extremely shallow depth of field, isolating the subject in its miniature world and crafting a serene and powerful atmosphere of natural art.", + "seed": 16255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_5/index_3167_seed_16255_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": -0.04577486119233072, + "vlm_logprob": -0.04577486119233072, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.04577486119233072, + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13204070925712585 + ], + "opposite_logprob_tokens": [ + -2.1320407390594482 + ], + "logprob_reward": -0.13204070925712585, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13204070925712585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029330749064683914, + 0.022489480674266815, + 0.03784333914518356, + 0.06950940191745758, + 0.4029419720172882, + 0.0422491692006588, + 0.030338646844029427, + 0.0386151522397995, + 0.07220464199781418, + 0.25447744131088257 + ], + "credit_map": [ + 0.0030076943803578615, + 0.004028474446386099, + 0.003259666496887803, + 0.004208068363368511, + 0.0035834680311381817, + 0.00328304385766387, + 0.004371476825326681, + 0.0035888596903532743, + 0.002360287820920348, + 0.002457018941640854, + 0.002626909641548991, + 0.0033948554191738367, + 0.0029698654543608427, + 0.002668259432539344, + 0.002681420184671879, + 0.0033308635465800762, + 0.00764338718727231, + 0.0029751022811979055, + 0.011353748850524426, + 0.005320494994521141, + 0.0018391942139714956, + 0.0020835211034864187, + 0.003352631349116564, + 0.003275256371125579, + 0.013854342512786388, + 0.008746498264372349, + 0.009681141935288906, + 0.008238094858825207, + 0.009442382492125034, + 0.00540198665112257, + 0.008014542981982231, + 0.006130417343229055, + 0.05125278979539871, + 0.06951484084129333, + 0.03977392986416817, + 0.045153066515922546, + 0.07828301191329956, + 0.025580940768122673, + 0.04782234504818916, + 0.04556103050708771, + 0.005674822721630335, + 0.004767828620970249, + 0.008027720265090466, + 0.006937549449503422, + 0.004137231968343258, + 0.0036378554068505764, + 0.0048710498958826065, + 0.00419511366635561, + 0.004632784519344568, + 0.002867514733225107, + 0.0037816260010004044, + 0.0033282607328146696, + 0.0042634205892682076, + 0.0024866261519491673, + 0.005724435206502676, + 0.0032539779786020517, + 0.003991320729255676, + 0.00272358488291502, + 0.007589838933199644, + 0.0036166496574878693, + 0.003840907709673047, + 0.00417078984901309, + 0.005754300858825445, + 0.0069277589209377766, + 0.007310716900974512, + 0.009056863375008106, + 0.013843556866049767, + 0.007647517137229443, + 0.013727844692766666, + 0.005143952090293169, + 0.009098075330257416, + 0.006376112811267376, + 0.056542061269283295, + 0.015906015411019325, + 0.08588477969169617, + 0.01804913766682148, + 0.018273109570145607, + 0.011526519432663918, + 0.025342918932437897, + 0.022952890023589134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.13204070925712585 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009802537970244884 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.009802537970244884, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009802537970244884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04515008628368378, + 0.05472981184720993, + 0.027699759230017662, + 0.079901322722435, + 0.19834966957569122, + 0.050751134753227234, + 0.080782450735569, + 0.07147243618965149, + 0.07123737037181854, + 0.3199259340763092 + ], + "credit_map": [ + 0.007100655231624842, + 0.006326872389763594, + 0.006741792894899845, + 0.006026884075254202, + 0.004632658790796995, + 0.003950415179133415, + 0.004749636631458998, + 0.00562116876244545, + 0.00816893856972456, + 0.004441431723535061, + 0.015122788958251476, + 0.004487507976591587, + 0.005068953149020672, + 0.0075698597356677055, + 0.004955724813044071, + 0.004914605990052223, + 0.0030196048319339752, + 0.0025255531072616577, + 0.003918625880032778, + 0.0032158188987523317, + 0.002419195370748639, + 0.0049207317642867565, + 0.0029422491788864136, + 0.004737981595098972, + 0.014695084653794765, + 0.014323955401778221, + 0.010381662286818027, + 0.009076237678527832, + 0.009329752065241337, + 0.005314049310982227, + 0.010547486133873463, + 0.006233104504644871, + 0.025464670732617378, + 0.03030194155871868, + 0.01834729313850403, + 0.03439412638545036, + 0.033237747848033905, + 0.010039429180324078, + 0.028225019574165344, + 0.018339429050683975, + 0.006588366348296404, + 0.006509811617434025, + 0.007632759399712086, + 0.0059594037011265755, + 0.004861489403992891, + 0.006574427243322134, + 0.006349016912281513, + 0.006275858264416456, + 0.011547173373401165, + 0.009689997881650925, + 0.010227363556623459, + 0.015737194567918777, + 0.007439909968525171, + 0.007755274884402752, + 0.01019185222685337, + 0.008193685673177242, + 0.014538899064064026, + 0.008455730974674225, + 0.008578390814363956, + 0.009102285839617252, + 0.010780883021652699, + 0.006461460608989, + 0.008204919286072254, + 0.005349862389266491, + 0.006966803222894669, + 0.009740444831550121, + 0.009528187103569508, + 0.008389678783714771, + 0.009104864671826363, + 0.007066281046718359, + 0.012320922687649727, + 0.008120188489556313, + 0.05563288927078247, + 0.03717819228768349, + 0.04751472920179367, + 0.03521301969885826, + 0.037545718252658844, + 0.02518489770591259, + 0.041809581220149994, + 0.03984692320227623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.009802537970244884 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002821395406499505 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.002821395406499505, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002821395406499505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0783018246293068, + 0.02252187766134739, + 0.011224632151424885, + 0.03300904482603073, + 0.3475814163684845, + 0.035425566136837006, + 0.051116202026605606, + 0.03663642331957817, + 0.17050059139728546, + 0.21368245780467987 + ], + "credit_map": [ + 0.0104842409491539, + 0.010371407493948936, + 0.012170038186013699, + 0.018164368346333504, + 0.009205058217048645, + 0.004516190849244595, + 0.00830074306577444, + 0.005089771468192339, + 0.0024760072119534016, + 0.0023137852549552917, + 0.0044545275159180164, + 0.0023616193793714046, + 0.0028149140998721123, + 0.0029614539816975594, + 0.0023529583122581244, + 0.0027866135351359844, + 0.0012538805603981018, + 0.0014706985093653202, + 0.0015351474285125732, + 0.001828556414693594, + 0.0015065098414197564, + 0.0009804946603253484, + 0.001638544606976211, + 0.0010107984999194741, + 0.006364478729665279, + 0.00589359924197197, + 0.0037854164838790894, + 0.004750088322907686, + 0.004122525453567505, + 0.002430277643725276, + 0.003039861563593149, + 0.0026227934285998344, + 0.04362788423895836, + 0.061665114015340805, + 0.035014908760786057, + 0.033406779170036316, + 0.06982292979955673, + 0.024104377254843712, + 0.041750017553567886, + 0.03818938881158829, + 0.005393030121922493, + 0.004796079825609922, + 0.005260785110294819, + 0.006407586392015219, + 0.0036570976953953505, + 0.001868381048552692, + 0.004870987962931395, + 0.0031716166995465755, + 0.004986352752894163, + 0.007567351683974266, + 0.005853865761309862, + 0.01201073918491602, + 0.0047421264462172985, + 0.0038755382411181927, + 0.007476107683032751, + 0.004604115150868893, + 0.00490971002727747, + 0.003816866083070636, + 0.0055089667439460754, + 0.0038175114896148443, + 0.005053787957876921, + 0.002621218329295516, + 0.005730161909013987, + 0.005178198684006929, + 0.015779059380292892, + 0.02621486783027649, + 0.023551730439066887, + 0.017585964873433113, + 0.03972022980451584, + 0.013846253044903278, + 0.019363263621926308, + 0.01443919911980629, + 0.04949239641427994, + 0.011168374679982662, + 0.07518963515758514, + 0.014230284839868546, + 0.013677699491381645, + 0.00955231487751007, + 0.020812788978219032, + 0.019558941945433617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.002821395406499505 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896333932876587 + ], + "opposite_logprob_tokens": [ + -2.578963279724121 + ], + "logprob_reward": -0.07896333932876587, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896333932876587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05691332742571831, + 0.01959092915058136, + 0.02343727834522724, + 0.08966860920190811, + 0.2677515149116516, + 0.06606721878051758, + 0.10127918422222137, + 0.05023245885968208, + 0.023257803171873093, + 0.3018016815185547 + ], + "credit_map": [ + 0.007504837121814489, + 0.006980752106755972, + 0.010120163671672344, + 0.008943015709519386, + 0.007651158142834902, + 0.0041657849214971066, + 0.0054355827160179615, + 0.006112034898251295, + 0.0019675155635923147, + 0.002758381888270378, + 0.002641728613525629, + 0.003399405162781477, + 0.002690058434382081, + 0.002054948825389147, + 0.0014650290831923485, + 0.0026138608809560537, + 0.0030870933551341295, + 0.0032582648564130068, + 0.0035871395375579596, + 0.0030131901148706675, + 0.003149650525301695, + 0.0021282886154949665, + 0.0021837761159986258, + 0.0030298742931336164, + 0.01702059991657734, + 0.014322880655527115, + 0.011065990664064884, + 0.014430719427764416, + 0.008731498382985592, + 0.00512792868539691, + 0.010317154228687286, + 0.008651839569211006, + 0.0349295400083065, + 0.0408826619386673, + 0.026846732944250107, + 0.0370870940387249, + 0.04789397865533829, + 0.01804550364613533, + 0.03252639248967171, + 0.029539616778492928, + 0.007450402714312077, + 0.00786410178989172, + 0.008229786530137062, + 0.00926289428025484, + 0.010164745151996613, + 0.005471247248351574, + 0.01021628174930811, + 0.007407763507217169, + 0.0114548709243536, + 0.009938470087945461, + 0.020305199548602104, + 0.025471489876508713, + 0.009627294726669788, + 0.004959234967827797, + 0.013260584324598312, + 0.0062620374374091625, + 0.0052842553704977036, + 0.004835008177906275, + 0.0063814930617809296, + 0.004913054872304201, + 0.007379703223705292, + 0.005437116604298353, + 0.00782787799835205, + 0.008173950016498566, + 0.0020357531029731035, + 0.003488080808892846, + 0.0032453443855047226, + 0.002486087614670396, + 0.005192662123590708, + 0.0017530869226902723, + 0.0031182626262307167, + 0.0019385246559977531, + 0.066990427672863, + 0.020505530759692192, + 0.09543582797050476, + 0.02267785184085369, + 0.02281879261136055, + 0.01529909111559391, + 0.02960527129471302, + 0.028468875214457512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.07896333932876587 + }, + { + "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 giant chalk beetle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the beetle positioned motionless on a vibrant green, dew-covered fern leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the beetle's carapace appear black with oil-slick iridescence showing deep green and purple hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three horns visible on the beetle's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, emerald blur creating a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246323999017477 + ], + "opposite_logprob_tokens": [ + -5.255246162414551 + ], + "logprob_reward": -0.005246323999017477, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246323999017477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12697196006774902, + 0.060129955410957336, + 0.04799211025238037, + 0.09125702828168869, + 0.11204955726861954, + 0.08902890980243683, + 0.17924194037914276, + 0.15085946023464203, + 0.02347426302731037, + 0.11899477243423462 + ], + "credit_map": [ + 0.01909978687763214, + 0.014423814602196217, + 0.032620999962091446, + 0.01619025319814682, + 0.010909473523497581, + 0.0088150380179286, + 0.013610485009849072, + 0.011302122846245766, + 0.00945205520838499, + 0.005305442027747631, + 0.016027871519327164, + 0.007448337972164154, + 0.0057406071573495865, + 0.004040129017084837, + 0.006083871237933636, + 0.006031648721545935, + 0.008124463260173798, + 0.004715061280876398, + 0.011732659302651882, + 0.008738024160265923, + 0.003342083655297756, + 0.002771727740764618, + 0.004264759831130505, + 0.004303331021219492, + 0.0197550468146801, + 0.01216970756649971, + 0.01011018455028534, + 0.013407927937805653, + 0.009999715723097324, + 0.005685380659997463, + 0.00961392279714346, + 0.010515157133340836, + 0.009495953097939491, + 0.010803651995956898, + 0.013310486450791359, + 0.03316846117377281, + 0.010264882817864418, + 0.004487043712288141, + 0.022397492080926895, + 0.008121595717966557, + 0.011243567802011967, + 0.011178449727594852, + 0.017990516498684883, + 0.007836352102458477, + 0.011340097524225712, + 0.00821282435208559, + 0.011578387580811977, + 0.00964872632175684, + 0.021615412086248398, + 0.02065412327647209, + 0.0268565583974123, + 0.0470501072704792, + 0.014954974874854088, + 0.009581100195646286, + 0.02588869072496891, + 0.012641001492738724, + 0.020664820447564125, + 0.016322704032063484, + 0.022526901215314865, + 0.020963599905371666, + 0.0232035294175148, + 0.011069091968238354, + 0.020991992205381393, + 0.015116848982870579, + 0.0026560903061181307, + 0.0027774828486144543, + 0.0024293058086186647, + 0.004337189253419638, + 0.003642205148935318, + 0.0018080465961247683, + 0.003814761294052005, + 0.0020091827027499676, + 0.022433511912822723, + 0.011794845573604107, + 0.02976200357079506, + 0.01022406481206417, + 0.01423096377402544, + 0.006820987910032272, + 0.012398811057209969, + 0.011329589411616325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.005246323999017477 + } + ], + "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/0000167/group_6/index_tensor([4636, 4636])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_tensor([4636, 4636])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9fa499556bb9c8112ecf8282ab677009ebfbd4ad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_tensor([4636, 4636])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, subjective point-of-view shot, shot from inside a car, reveals a side window covered in dense raindrops. Each droplet acts like a miniature lens, distorting and refracting the world outside. The raindrops slowly slide down the glass, leaving behind blurred streaks that add a hazy, tense quality to the scene. Outside, a wet city alley stretches out into the night. The asphalt pavement reflects the flickering red, blue, and green neon lights of the surrounding shop signs, creating splatters of oscillating watercolor. Amidst this dazzling yet chilling light, two men stand face to face: one is tall, clad in a long, dark trench coat, its soaked fabric clinging to his tense shoulders. He leans slightly forward, his posture imbued with oppression, his facial expression sullen yet resolute in the shifting light. The other, younger, wears a black hoodie, his face half hidden in shadow, half illuminated by a blue neon glow. His arms are slightly outstretched, his hands clenched into half fists, his eyes a mixture of vigilance and unyielding determination. They stand a few steps apart, motionless, yet the air is thick with tension, as if a confrontation could erupt at any moment. The entire scene, shrouded in cold tones and high-contrast lighting, creates a cinematic experience of profound suspense and uncertainty.", + "seed": 3706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3706_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": -0.018719751760363577, + "vlm_logprob": -0.018719751760363577, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.018719751760363577, + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02287360467016697 + ], + "opposite_logprob_tokens": [ + -4.272873401641846 + ], + "logprob_reward": -0.02287360467016697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02287360467016697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24482464790344238, + 0.18374890089035034, + 0.046287283301353455, + 0.13084596395492554, + 0.1262284219264984, + 0.045670684427022934, + 0.048028092831373215, + 0.047989360988140106, + 0.07319214940071106, + 0.05318450555205345 + ], + "credit_map": [ + 0.02700068987905979, + 0.02707645669579506, + 0.03734064847230911, + 0.03293401747941971, + 0.03279075026512146, + 0.023589322343468666, + 0.041568100452423096, + 0.0225246362388134, + 0.013659581542015076, + 0.027628786861896515, + 0.016696391627192497, + 0.029211856424808502, + 0.022430552169680595, + 0.02143375389277935, + 0.028072725981473923, + 0.024615228176116943, + 0.005355037283152342, + 0.005757730919867754, + 0.005131998565047979, + 0.00759749673306942, + 0.004568811040371656, + 0.004122383426874876, + 0.006070464383810759, + 0.007683358620852232, + 0.012059926055371761, + 0.01082345936447382, + 0.013956526294350624, + 0.020154837518930435, + 0.009947932325303555, + 0.019695846363902092, + 0.02248605526983738, + 0.021721376106142998, + 0.013480656780302525, + 0.010961807332932949, + 0.016909990459680557, + 0.01936883106827736, + 0.012083579786121845, + 0.015560694970190525, + 0.016404764726758003, + 0.021458083763718605, + 0.005911585409194231, + 0.003762563457712531, + 0.0069192335940897465, + 0.00875748973339796, + 0.003013261128216982, + 0.004489912185817957, + 0.0061883204616606236, + 0.006628310773521662, + 0.007002581842243671, + 0.005242811515927315, + 0.006083170883357525, + 0.007397481240332127, + 0.003632782958447933, + 0.0041635241359472275, + 0.00614249799400568, + 0.008363236673176289, + 0.006205527111887932, + 0.0031435645651072264, + 0.009571249596774578, + 0.011222820729017258, + 0.002727619605138898, + 0.004079971928149462, + 0.005640712101012468, + 0.005397884640842676, + 0.012035487219691277, + 0.008552835322916508, + 0.009861404076218605, + 0.008431527763605118, + 0.00835000816732645, + 0.008296775631606579, + 0.009391391649842262, + 0.008272715844213963, + 0.0047140344977378845, + 0.005622750148177147, + 0.006034454796463251, + 0.007432118058204651, + 0.005947658326476812, + 0.007077896036207676, + 0.007803223095834255, + 0.00855235941708088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02287360467016697 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801357984542847 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.009801357984542847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801357984542847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12678827345371246, + 0.11966809630393982, + 0.023391827940940857, + 0.15091440081596375, + 0.14654302597045898, + 0.056496620178222656, + 0.035955242812633514, + 0.08862607926130295, + 0.1700596958398819, + 0.08155682682991028 + ], + "credit_map": [ + 0.01652994006872177, + 0.01597864180803299, + 0.016341667622327805, + 0.016266997903585434, + 0.019397277384996414, + 0.011018547229468822, + 0.01928287371993065, + 0.011972315609455109, + 0.012064574286341667, + 0.013715392909944057, + 0.016034509986639023, + 0.019158214330673218, + 0.01374438963830471, + 0.012177957221865654, + 0.019554397091269493, + 0.013218644075095654, + 0.0031128041446208954, + 0.002029755851253867, + 0.00240358361043036, + 0.003790745045989752, + 0.0016818351577967405, + 0.0034462781623005867, + 0.0037756538949906826, + 0.003151169279590249, + 0.014029075391590595, + 0.011524230241775513, + 0.01601928099989891, + 0.022745715454220772, + 0.012807856313884258, + 0.02443469502031803, + 0.027558794245123863, + 0.021794741973280907, + 0.010806401260197163, + 0.009168489836156368, + 0.015135594643652439, + 0.022991085425019264, + 0.010535528883337975, + 0.01737036369740963, + 0.021358095109462738, + 0.039177436381578445, + 0.004767653066664934, + 0.00534897530451417, + 0.007555002346634865, + 0.007984848693013191, + 0.005385546013712883, + 0.008469829335808754, + 0.007600931916385889, + 0.00938382651656866, + 0.0042170751839876175, + 0.003019690979272127, + 0.005655074026435614, + 0.005235391203314066, + 0.002537920605391264, + 0.0041263652965426445, + 0.005840402562171221, + 0.005323317833244801, + 0.00696868309751153, + 0.005162022542208433, + 0.018191052600741386, + 0.022089794278144836, + 0.0062231686897575855, + 0.007154784165322781, + 0.013308048248291016, + 0.00952851865440607, + 0.013866309076547623, + 0.01616973988711834, + 0.025484859943389893, + 0.026736391708254814, + 0.016841650009155273, + 0.020012320950627327, + 0.029037246480584145, + 0.02191115729510784, + 0.009756671264767647, + 0.008889425545930862, + 0.009302930906414986, + 0.008864996023476124, + 0.009983641095459461, + 0.009635582566261292, + 0.012678545899689198, + 0.012445022352039814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.009801357984542847 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645626526325941 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645626526325941, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645626526325941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09652811288833618, + 0.10934051126241684, + 0.05273592844605446, + 0.23657779395580292, + 0.095203697681427, + 0.0670623779296875, + 0.05470512807369232, + 0.11711451411247253, + 0.06700185686349869, + 0.10373006761074066 + ], + "credit_map": [ + 0.015973873436450958, + 0.014873996376991272, + 0.014918074011802673, + 0.012412047013640404, + 0.011400860734283924, + 0.006006911862641573, + 0.01287711039185524, + 0.008065243251621723, + 0.010389856062829494, + 0.014441832900047302, + 0.012808740139007568, + 0.017032017931342125, + 0.012961720116436481, + 0.013247947208583355, + 0.014408192597329617, + 0.014050212688744068, + 0.005107089877128601, + 0.004774948116391897, + 0.005691925063729286, + 0.007676247041672468, + 0.0039190431125462055, + 0.0053636375814676285, + 0.007659936789423227, + 0.012543099001049995, + 0.026004916056990623, + 0.021151557564735413, + 0.022505227476358414, + 0.03406623750925064, + 0.01758994720876217, + 0.039181679487228394, + 0.04158071056008339, + 0.03449751064181328, + 0.009751790203154087, + 0.006947703659534454, + 0.014624137431383133, + 0.014092031866312027, + 0.007888569496572018, + 0.01069796085357666, + 0.01672380045056343, + 0.014477708376944065, + 0.010901179164648056, + 0.006478994619101286, + 0.011595710180699825, + 0.00885864906013012, + 0.005345786456018686, + 0.00747634656727314, + 0.008556333370506763, + 0.007849377579987049, + 0.006991734262555838, + 0.004858905915170908, + 0.0096627501770854, + 0.007410590536892414, + 0.0036186911165714264, + 0.005355723202228546, + 0.007331927306950092, + 0.009474802762269974, + 0.009622069075703621, + 0.008471856825053692, + 0.024605944752693176, + 0.03296243026852608, + 0.007115065585821867, + 0.008801136165857315, + 0.013011668808758259, + 0.012524341233074665, + 0.006981923244893551, + 0.00635260995477438, + 0.010778963565826416, + 0.007683307398110628, + 0.005435131024569273, + 0.005608915816992521, + 0.009852432645857334, + 0.014308573678135872, + 0.011918803676962852, + 0.008202958852052689, + 0.012897465378046036, + 0.011201628483831882, + 0.01090856920927763, + 0.011826988309621811, + 0.01891285739839077, + 0.01786080189049244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.004645626526325941 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13592807948589325, + 0.11655738949775696, + 0.07235541194677353, + 0.15076301991939545, + 0.0969807356595993, + 0.12058805674314499, + 0.03641655668616295, + 0.07729561626911163, + 0.0728164091706276, + 0.12029870599508286 + ], + "credit_map": [ + 0.024383826181292534, + 0.020945105701684952, + 0.021811190992593765, + 0.01751464046537876, + 0.018000999465584755, + 0.0072306077927351, + 0.017058830708265305, + 0.00898287259042263, + 0.013716683723032475, + 0.015300551429390907, + 0.01527322456240654, + 0.017391255125403404, + 0.013626971282064915, + 0.012323150411248207, + 0.016263188794255257, + 0.012662360444664955, + 0.006305493414402008, + 0.005812754388898611, + 0.006791536696255207, + 0.010555438697338104, + 0.004767068661749363, + 0.010131031274795532, + 0.010952144861221313, + 0.017039941623806953, + 0.02332429587841034, + 0.012912530452013016, + 0.018534451723098755, + 0.022663651034235954, + 0.010521700605750084, + 0.02544238604605198, + 0.01719902828335762, + 0.0201649758964777, + 0.009818905033171177, + 0.006513852626085281, + 0.013471955433487892, + 0.014118104241788387, + 0.007829666137695312, + 0.01405384298413992, + 0.014160802587866783, + 0.017013616859912872, + 0.008651873096823692, + 0.011147168464958668, + 0.013889383524656296, + 0.020422697067260742, + 0.011302454397082329, + 0.015589555725455284, + 0.01645178347826004, + 0.02313315123319626, + 0.004459585063159466, + 0.0031566207762807608, + 0.006344734225422144, + 0.005597901530563831, + 0.002547656651586294, + 0.003934971988201141, + 0.005044607911258936, + 0.005330480169504881, + 0.008405279368162155, + 0.006479271687567234, + 0.014387113973498344, + 0.019796937704086304, + 0.00373455579392612, + 0.006047572009265423, + 0.008195399306714535, + 0.010249488987028599, + 0.007799418643116951, + 0.007726428098976612, + 0.011918249540030956, + 0.0077890497632324696, + 0.005516708828508854, + 0.007986459881067276, + 0.009505309164524078, + 0.014574780128896236, + 0.012776573188602924, + 0.008917738683521748, + 0.0164223350584507, + 0.014147689566016197, + 0.01229067798703909, + 0.01206246018409729, + 0.022907091304659843, + 0.02077413722872734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.007626228500157595 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865194112062454 + ], + "opposite_logprob_tokens": [ + -3.048651933670044 + ], + "logprob_reward": -0.04865194112062454, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865194112062454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1480664759874344, + 0.06093493476510048, + 0.039436232298612595, + 0.12592679262161255, + 0.18055611848831177, + 0.10099508613348007, + 0.07439212501049042, + 0.0819762572646141, + 0.09608878940343857, + 0.09162717312574387 + ], + "credit_map": [ + 0.022074175998568535, + 0.022229021415114403, + 0.023135848343372345, + 0.0192424263805151, + 0.02000272646546364, + 0.011510536074638367, + 0.017810694873332977, + 0.012061057612299919, + 0.007332392502576113, + 0.004887662827968597, + 0.008219253271818161, + 0.010754427872598171, + 0.005286737810820341, + 0.00733543187379837, + 0.01025572419166565, + 0.006863309070467949, + 0.004196507390588522, + 0.003844362683594227, + 0.004706619773060083, + 0.005345745477825403, + 0.003908707294613123, + 0.0039551411755383015, + 0.005182219669222832, + 0.008296927437186241, + 0.012258174829185009, + 0.00987360067665577, + 0.015449888072907925, + 0.018764672800898552, + 0.010161362588405609, + 0.02169635146856308, + 0.013721923343837261, + 0.024000823497772217, + 0.013772623613476753, + 0.012159881182014942, + 0.0190286748111248, + 0.02775106020271778, + 0.008527107536792755, + 0.01796160638332367, + 0.03132586181163788, + 0.050029315054416656, + 0.010157373733818531, + 0.011733772233128548, + 0.012303005903959274, + 0.019130947068333626, + 0.010495124384760857, + 0.011640181764960289, + 0.013207877054810524, + 0.012326815165579319, + 0.00968414917588234, + 0.005491714924573898, + 0.012588218785822392, + 0.010583613067865372, + 0.004157237708568573, + 0.006506036035716534, + 0.011138481087982655, + 0.014242683537304401, + 0.00865684449672699, + 0.007977528497576714, + 0.011948998086154461, + 0.011636571027338505, + 0.005847162567079067, + 0.010040710680186749, + 0.013543233275413513, + 0.012325216084718704, + 0.0071552651934325695, + 0.007376680616289377, + 0.012514094822108746, + 0.016966819763183594, + 0.010224697180092335, + 0.0128614017739892, + 0.018046503886580467, + 0.010943337343633175, + 0.009028207510709763, + 0.005745221395045519, + 0.01137524377554655, + 0.009388642385601997, + 0.009609974920749664, + 0.012305631302297115, + 0.019590148702263832, + 0.014584106393158436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.04865194112062454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, subjective point-of-view shot, shot from inside a car, reveals a side window covered in dense raindrops. Each droplet acts like a miniature lens, distorting and refracting the world outside. The raindrops slowly slide down the glass, leaving behind blurred streaks that add a hazy, tense quality to the scene. Outside, a wet city alley stretches out into the night. The asphalt pavement reflects the flickering red, blue, and green neon lights of the surrounding shop signs, creating splatters of oscillating watercolor. Amidst this dazzling yet chilling light, two men stand face to face: one is tall, clad in a long, dark trench coat, its soaked fabric clinging to his tense shoulders. He leans slightly forward, his posture imbued with oppression, his facial expression sullen yet resolute in the shifting light. The other, younger, wears a black hoodie, his face half hidden in shadow, half illuminated by a blue neon glow. His arms are slightly outstretched, his hands clenched into half fists, his eyes a mixture of vigilance and unyielding determination. They stand a few steps apart, motionless, yet the air is thick with tension, as if a confrontation could erupt at any moment. The entire scene, shrouded in cold tones and high-contrast lighting, creates a cinematic experience of profound suspense and uncertainty.", + "seed": 3707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3707_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": -0.03712175544351339, + "vlm_logprob": -0.03712175544351339, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03712175544351339, + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022371888160705566 + ], + "opposite_logprob_tokens": [ + -4.647371768951416 + ], + "logprob_reward": -0.022371888160705566, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.022371888160705566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2912237346172333, + 0.1967461258172989, + 0.0676008015871048, + 0.11086716502904892, + 0.10006344318389893, + 0.048775508999824524, + 0.054340116679668427, + 0.04615290090441704, + 0.034093618392944336, + 0.05013655498623848 + ], + "credit_map": [ + 0.03704192116856575, + 0.04604875296354294, + 0.0549054890871048, + 0.03813367336988449, + 0.023461593315005302, + 0.029878946021199226, + 0.026569250971078873, + 0.035184118896722794, + 0.02162664756178856, + 0.012758146040141582, + 0.026210477575659752, + 0.03847142308950424, + 0.010980595834553242, + 0.04375071078538895, + 0.01941840536892414, + 0.02352975495159626, + 0.007240590639412403, + 0.004882331471890211, + 0.013465429656207561, + 0.006741471588611603, + 0.004322894848883152, + 0.01590282842516899, + 0.006179797928780317, + 0.00886546354740858, + 0.008936716243624687, + 0.01196968276053667, + 0.01041217241436243, + 0.012451980262994766, + 0.008733753114938736, + 0.02095014601945877, + 0.02231578528881073, + 0.015096943825483322, + 0.008247329853475094, + 0.009605498984456062, + 0.01743893511593342, + 0.017260877415537834, + 0.008606510236859322, + 0.010833179578185081, + 0.01255725510418415, + 0.015513871796429157, + 0.004572788253426552, + 0.004951043054461479, + 0.006802867632359266, + 0.0057251774705946445, + 0.003546491963788867, + 0.0073267268016934395, + 0.008085865527391434, + 0.007764554116874933, + 0.0067082978785037994, + 0.003920688759535551, + 0.008242766372859478, + 0.007512781303375959, + 0.0038289569783955812, + 0.0073856995441019535, + 0.00690228259190917, + 0.009838646277785301, + 0.005779725965112448, + 0.004235295113176107, + 0.005904827732592821, + 0.007565940264612436, + 0.0030273646116256714, + 0.006346320267766714, + 0.00556702958419919, + 0.007726404815912247, + 0.0052000717259943485, + 0.0038716101553291082, + 0.004976239986717701, + 0.004979120101779699, + 0.0020989063195884228, + 0.0028413671534508467, + 0.004182362463325262, + 0.005943941418081522, + 0.008420265279710293, + 0.005390894133597612, + 0.009911767207086086, + 0.004432324320077896, + 0.0033901692368090153, + 0.005792280659079552, + 0.006228331942111254, + 0.006570528261363506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.022371888160705566 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095019057393074 + ], + "opposite_logprob_tokens": [ + -4.51109504699707 + ], + "logprob_reward": -0.011095019057393074, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095019057393074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07635585218667984, + 0.36128169298171997, + 0.06286504864692688, + 0.026977339759469032, + 0.13733424246311188, + 0.08053469657897949, + 0.09032168239355087, + 0.0784783661365509, + 0.04303766414523125, + 0.04281338304281235 + ], + "credit_map": [ + 0.009567390196025372, + 0.006993983872234821, + 0.013103190809488297, + 0.010444934479892254, + 0.005367751698940992, + 0.01280269492417574, + 0.007034318521618843, + 0.011041587218642235, + 0.042616117745637894, + 0.023350676521658897, + 0.03233295679092407, + 0.08596920222043991, + 0.017861267551779747, + 0.08327946811914444, + 0.03520343452692032, + 0.0406685508787632, + 0.00526023143902421, + 0.004395743366330862, + 0.01636088453233242, + 0.005606363527476788, + 0.0035831895656883717, + 0.018750283867120743, + 0.004070327617228031, + 0.004838025197386742, + 0.0029909766744822264, + 0.0034202493261545897, + 0.0036575966514647007, + 0.0030451614875346422, + 0.002690225141122937, + 0.0044055418111383915, + 0.00379050406627357, + 0.0029770852997899055, + 0.010167605243623257, + 0.006184330675750971, + 0.04240070655941963, + 0.013632376678287983, + 0.005404271651059389, + 0.009515161626040936, + 0.009681605733931065, + 0.04034819081425667, + 0.005221622064709663, + 0.007148665841668844, + 0.008127298206090927, + 0.007170373573899269, + 0.007771275006234646, + 0.01979970373213291, + 0.01006142795085907, + 0.015234330669045448, + 0.008557007648050785, + 0.009020316414535046, + 0.010950666852295399, + 0.013164039701223373, + 0.005468250717967749, + 0.018594183027744293, + 0.008574198000133038, + 0.015993019565939903, + 0.008873241022229195, + 0.009419077076017857, + 0.007499499246478081, + 0.008727294392883778, + 0.008181369863450527, + 0.0125883799046278, + 0.010210508480668068, + 0.012978995218873024, + 0.004308787174522877, + 0.003698283340781927, + 0.006739652715623379, + 0.007000353652983904, + 0.0025815595872700214, + 0.008628571406006813, + 0.004385425243526697, + 0.005695032887160778, + 0.0031071940902620554, + 0.0036196960136294365, + 0.005065260920673609, + 0.0035973661579191685, + 0.007974772714078426, + 0.004867998883128166, + 0.008874254301190376, + 0.005706841126084328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.011095019057393074 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546890415251255 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012546890415251255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546890415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07284693419933319, + 0.08982276171445847, + 0.2103816419839859, + 0.06068245321512222, + 0.05830228701233864, + 0.1079518049955368, + 0.08530007302761078, + 0.10236001759767532, + 0.16171196103096008, + 0.0506400540471077 + ], + "credit_map": [ + 0.008205651305615902, + 0.012163268402218819, + 0.008941035717725754, + 0.012132315896451473, + 0.005393805913627148, + 0.0061583444476127625, + 0.005705376621335745, + 0.014147136360406876, + 0.00972836185246706, + 0.00758814811706543, + 0.009923025034368038, + 0.023586902767419815, + 0.004423118662089109, + 0.017087794840335846, + 0.008215720765292645, + 0.009269683621823788, + 0.015164765529334545, + 0.00714996038004756, + 0.0630582943558693, + 0.0184823889285326, + 0.008430229499936104, + 0.07123735547065735, + 0.011738980188965797, + 0.015119673684239388, + 0.004905964247882366, + 0.0070875464007258415, + 0.0064691039733588696, + 0.007320405449718237, + 0.007408495061099529, + 0.009555303491652012, + 0.007322050165385008, + 0.010613585822284222, + 0.00477595767006278, + 0.005744426976889372, + 0.00798141397535801, + 0.008904330432415009, + 0.005573485977947712, + 0.00841338001191616, + 0.009404558688402176, + 0.00750473327934742, + 0.009529376402497292, + 0.0075292219407856464, + 0.02037820778787136, + 0.009811781346797943, + 0.006207723170518875, + 0.025540050119161606, + 0.011955713853240013, + 0.016999725252389908, + 0.009684859775006771, + 0.008339310064911842, + 0.010768851265311241, + 0.013163790106773376, + 0.006123557686805725, + 0.015026725828647614, + 0.008417367935180664, + 0.013775608502328396, + 0.02268289215862751, + 0.010558836162090302, + 0.013674120418727398, + 0.012610960751771927, + 0.004522029310464859, + 0.012258569709956646, + 0.009419823996722698, + 0.016632787883281708, + 0.027066413313150406, + 0.013211110606789589, + 0.028302984312176704, + 0.027209166437387466, + 0.007511053700000048, + 0.015603910200297832, + 0.014512035064399242, + 0.02829528972506523, + 0.005869285203516483, + 0.005576109979301691, + 0.005784605164080858, + 0.0042082988657057285, + 0.005351464729756117, + 0.00917577650398016, + 0.007464527618139982, + 0.007209988310933113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.012546890415251255 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537471950054169 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537471950054169, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537471950054169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08378773927688599, + 0.07527346909046173, + 0.06886500865221024, + 0.05281895771622658, + 0.2036137729883194, + 0.1874566227197647, + 0.09294552356004715, + 0.07623742520809174, + 0.05809491500258446, + 0.100906603038311 + ], + "credit_map": [ + 0.007894802838563919, + 0.011073979549109936, + 0.0101932967081666, + 0.014842595905065536, + 0.006071827840059996, + 0.006777073722332716, + 0.008776216767728329, + 0.01815793290734291, + 0.006982671562582254, + 0.006345085799694061, + 0.014507601037621498, + 0.016331680119037628, + 0.005268350709229708, + 0.009485832415521145, + 0.007931249216198921, + 0.008420992642641068, + 0.008385389111936092, + 0.006355311255902052, + 0.009496521204710007, + 0.007543409243226051, + 0.006270735524594784, + 0.0129076624289155, + 0.008483371697366238, + 0.0094226049259305, + 0.005588275380432606, + 0.006266871467232704, + 0.006542954593896866, + 0.006429639644920826, + 0.0065365200862288475, + 0.006866981275379658, + 0.0072256457060575485, + 0.007362060714513063, + 0.013845489360392094, + 0.015291242860257626, + 0.04556608945131302, + 0.02784397266805172, + 0.015162748284637928, + 0.02307118847966194, + 0.02084619738161564, + 0.041986823081970215, + 0.023599514737725258, + 0.012627067044377327, + 0.03283223509788513, + 0.018637971952557564, + 0.013129863888025284, + 0.03345219045877457, + 0.0238614771515131, + 0.029316294938325882, + 0.008769407868385315, + 0.009084873832762241, + 0.00981979351490736, + 0.015501867979764938, + 0.007557723205536604, + 0.015203605405986309, + 0.014193743467330933, + 0.012814490124583244, + 0.008658075705170631, + 0.0063406480476260185, + 0.010223516263067722, + 0.011445612646639347, + 0.0046701328828930855, + 0.012113356962800026, + 0.011537311598658562, + 0.011248761788010597, + 0.0063175936229527, + 0.0036031901836395264, + 0.01093317847698927, + 0.011537709273397923, + 0.003758567851036787, + 0.006119892932474613, + 0.006609494797885418, + 0.009215281344950199, + 0.0059007215313613415, + 0.010141340084373951, + 0.007806792855262756, + 0.010003665462136269, + 0.010010884143412113, + 0.02486853301525116, + 0.01511099748313427, + 0.017063651233911514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.012537471950054169 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270575076341629 + ], + "opposite_logprob_tokens": [ + -2.1270575523376465 + ], + "logprob_reward": -0.1270575076341629, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270575076341629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10411040484905243, + 0.16507399082183838, + 0.08156831562519073, + 0.11575674265623093, + 0.12665912508964539, + 0.10650669783353806, + 0.10459979623556137, + 0.04543619602918625, + 0.051392700523138046, + 0.09889598190784454 + ], + "credit_map": [ + 0.01485133171081543, + 0.0220792293548584, + 0.014722156338393688, + 0.01249409094452858, + 0.009281794540584087, + 0.010673757642507553, + 0.009653055109083652, + 0.010354995727539062, + 0.02036006562411785, + 0.010612351819872856, + 0.01684056781232357, + 0.037711456418037415, + 0.008232097141444683, + 0.03652937337756157, + 0.016351565718650818, + 0.01843651756644249, + 0.00934692844748497, + 0.006177516188472509, + 0.019602445885539055, + 0.007384823635220528, + 0.005844203755259514, + 0.020272213965654373, + 0.0066968416795134544, + 0.00624334579333663, + 0.010872987098991871, + 0.010839373804628849, + 0.018608946353197098, + 0.013378867879509926, + 0.012111417949199677, + 0.01836991123855114, + 0.018635373562574387, + 0.012939867563545704, + 0.007623518351465464, + 0.008206439204514027, + 0.028832411393523216, + 0.011897914111614227, + 0.010607822798192501, + 0.016354713588953018, + 0.016612164676189423, + 0.02652413584291935, + 0.010347386822104454, + 0.007647502701729536, + 0.011855922639369965, + 0.006228198297321796, + 0.013013161718845367, + 0.023682229220867157, + 0.01571967825293541, + 0.018012618646025658, + 0.012345138937234879, + 0.008412616327404976, + 0.01440427452325821, + 0.01306475605815649, + 0.007704662159085274, + 0.014815018512308598, + 0.016749082133173943, + 0.0171042513102293, + 0.0041990275494754314, + 0.003985884599387646, + 0.006046985276043415, + 0.006817222107201815, + 0.0033822546247392893, + 0.004918535239994526, + 0.009556982666254044, + 0.006529304198920727, + 0.006602204870432615, + 0.004115737974643707, + 0.008843351155519485, + 0.006776613648980856, + 0.005192547105252743, + 0.006866724230349064, + 0.006471679545938969, + 0.006523840595036745, + 0.009957759641110897, + 0.009769199416041374, + 0.011738648638129234, + 0.010386425070464611, + 0.009286702610552311, + 0.02025279588997364, + 0.012417718768119812, + 0.015086732804775238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.1270575076341629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, subjective point-of-view shot, shot from inside a car, reveals a side window covered in dense raindrops. Each droplet acts like a miniature lens, distorting and refracting the world outside. The raindrops slowly slide down the glass, leaving behind blurred streaks that add a hazy, tense quality to the scene. Outside, a wet city alley stretches out into the night. The asphalt pavement reflects the flickering red, blue, and green neon lights of the surrounding shop signs, creating splatters of oscillating watercolor. Amidst this dazzling yet chilling light, two men stand face to face: one is tall, clad in a long, dark trench coat, its soaked fabric clinging to his tense shoulders. He leans slightly forward, his posture imbued with oppression, his facial expression sullen yet resolute in the shifting light. The other, younger, wears a black hoodie, his face half hidden in shadow, half illuminated by a blue neon glow. His arms are slightly outstretched, his hands clenched into half fists, his eyes a mixture of vigilance and unyielding determination. They stand a few steps apart, motionless, yet the air is thick with tension, as if a confrontation could erupt at any moment. The entire scene, shrouded in cold tones and high-contrast lighting, creates a cinematic experience of profound suspense and uncertainty.", + "seed": 3708, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3708_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": -0.04012237144634127, + "vlm_logprob": -0.04012237144634127, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04012237144634127, + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040923118591308594 + ], + "opposite_logprob_tokens": [ + -3.5409231185913086 + ], + "logprob_reward": -0.040923118591308594, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.040923118591308594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18431052565574646, + 0.11078663170337677, + 0.0672951340675354, + 0.12504644691944122, + 0.10218252241611481, + 0.11089389026165009, + 0.04927736520767212, + 0.11483565717935562, + 0.06112387403845787, + 0.07424800097942352 + ], + "credit_map": [ + 0.02598702348768711, + 0.027153095230460167, + 0.0268598273396492, + 0.02164980210363865, + 0.016169674694538116, + 0.012527886778116226, + 0.02856694906949997, + 0.025396263226866722, + 0.01035420224070549, + 0.008992061950266361, + 0.012170876376330853, + 0.02082807570695877, + 0.011740957386791706, + 0.008526746183633804, + 0.024165255948901176, + 0.01400846429169178, + 0.005181171000003815, + 0.006357898935675621, + 0.009451876394450665, + 0.006192563101649284, + 0.007111010607331991, + 0.01043553464114666, + 0.010397774167358875, + 0.012167306616902351, + 0.009475491009652615, + 0.015180792659521103, + 0.011535121127963066, + 0.018370574340224266, + 0.021287374198436737, + 0.0116432448849082, + 0.016860079020261765, + 0.020693767815828323, + 0.008413903415203094, + 0.015284841880202293, + 0.00783549528568983, + 0.017360582947731018, + 0.015363173559308052, + 0.011629789136350155, + 0.015441673807799816, + 0.010853055864572525, + 0.008953518234193325, + 0.018368927761912346, + 0.009298630058765411, + 0.02393004484474659, + 0.006578208412975073, + 0.010289889760315418, + 0.008580628782510757, + 0.024894049391150475, + 0.00541708804666996, + 0.0037859962321817875, + 0.006432269234210253, + 0.006809463258832693, + 0.00533171184360981, + 0.007938706316053867, + 0.005507623311132193, + 0.008054504171013832, + 0.014766091480851173, + 0.006628985051065683, + 0.009857448749244213, + 0.02706083655357361, + 0.008086557500064373, + 0.009150728583335876, + 0.007600503508001566, + 0.0316845066845417, + 0.006319162901490927, + 0.010264106094837189, + 0.006050998345017433, + 0.01123454887419939, + 0.0045414515770971775, + 0.006911704782396555, + 0.005204527173191309, + 0.01059737242758274, + 0.00720467371866107, + 0.009235309436917305, + 0.007078984286636114, + 0.01129644550383091, + 0.013377017341554165, + 0.009493918158113956, + 0.007615160662680864, + 0.008946491405367851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.040923118591308594 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016100509092211723 + ], + "opposite_logprob_tokens": [ + -4.1411004066467285 + ], + "logprob_reward": -0.016100509092211723, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016100509092211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1070530116558075, + 0.05222035571932793, + 0.02828231453895569, + 0.13851676881313324, + 0.0654669925570488, + 0.2819893956184387, + 0.10835614800453186, + 0.10219557583332062, + 0.0455755740404129, + 0.07034384459257126 + ], + "credit_map": [ + 0.015969261527061462, + 0.012890923768281937, + 0.015264902263879776, + 0.011596261523663998, + 0.012203075923025608, + 0.011012591421604156, + 0.015863429754972458, + 0.01225257758051157, + 0.004832439590245485, + 0.004770898725837469, + 0.006681291852146387, + 0.008012776263058186, + 0.005386644043028355, + 0.004385385662317276, + 0.01168100070208311, + 0.006469921674579382, + 0.002946011256426573, + 0.0037756608799099922, + 0.004570527467876673, + 0.0029184583108872175, + 0.002408271422609687, + 0.0034314522054046392, + 0.00415199575945735, + 0.00407993933185935, + 0.012365472503006458, + 0.027594702318310738, + 0.014850998297333717, + 0.01933618076145649, + 0.02286345697939396, + 0.012197459116578102, + 0.014668564312160015, + 0.014639947563409805, + 0.007858253084123135, + 0.011461462825536728, + 0.006183631718158722, + 0.009646047838032246, + 0.007151271216571331, + 0.005609525367617607, + 0.01096829678863287, + 0.006588509306311607, + 0.01816229149699211, + 0.035139262676239014, + 0.02176739275455475, + 0.049072954803705215, + 0.023768674582242966, + 0.06467951089143753, + 0.03023257479071617, + 0.03916672617197037, + 0.012891534715890884, + 0.011323293671011925, + 0.009984527714550495, + 0.017816755920648575, + 0.012107422575354576, + 0.011974023655056953, + 0.01857239380478859, + 0.01368620153516531, + 0.005914533510804176, + 0.008608912117779255, + 0.01042853482067585, + 0.02834072709083557, + 0.004662540275603533, + 0.007333275396376848, + 0.0050462475046515465, + 0.03186081722378731, + 0.004973323550075293, + 0.0061025312170386314, + 0.004493787419050932, + 0.008290420286357403, + 0.005343212746083736, + 0.0038156432565301657, + 0.004912877455353737, + 0.007643779739737511, + 0.0057641929015517235, + 0.009753799065947533, + 0.005859841592609882, + 0.012731023132801056, + 0.009491287171840668, + 0.008169366978108883, + 0.005751473363488913, + 0.01282286923378706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.016100509092211723 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00675099017098546 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.00675099017098546, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00675099017098546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07520216703414917, + 0.06072184443473816, + 0.03973443806171417, + 0.05753064155578613, + 0.10013797134160995, + 0.1119033694267273, + 0.09310979396104813, + 0.30204206705093384, + 0.08447012305259705, + 0.0751475915312767 + ], + "credit_map": [ + 0.006569994613528252, + 0.009615588933229446, + 0.010894903913140297, + 0.01449241116642952, + 0.008586794137954712, + 0.009741371497511864, + 0.008044067770242691, + 0.007257028482854366, + 0.006638023070991039, + 0.005520206410437822, + 0.007031942717730999, + 0.010253891348838806, + 0.006759805604815483, + 0.004841145593672991, + 0.012784099206328392, + 0.00689272303134203, + 0.004493741784244776, + 0.0049903495237231255, + 0.005717027001082897, + 0.004141008947044611, + 0.0036343098618090153, + 0.003903367090970278, + 0.0077388775534927845, + 0.005115755368024111, + 0.00556157436221838, + 0.0073992665857076645, + 0.006736183539032936, + 0.007079731673002243, + 0.007143842987716198, + 0.005111634731292725, + 0.00789441354572773, + 0.010603993199765682, + 0.007472294848412275, + 0.011862121522426605, + 0.009382515214383602, + 0.016089193522930145, + 0.008504778146743774, + 0.012879833579063416, + 0.02254561334848404, + 0.011401616968214512, + 0.009516906924545765, + 0.017548471689224243, + 0.014090270735323429, + 0.02068362385034561, + 0.008754056878387928, + 0.011928201653063297, + 0.012504625134170055, + 0.016877207905054092, + 0.012933794409036636, + 0.011495593003928661, + 0.0074815344996750355, + 0.010711442679166794, + 0.011407977901399136, + 0.015294797718524933, + 0.010509883984923363, + 0.013274758122861385, + 0.018951795995235443, + 0.05038023367524147, + 0.030538424849510193, + 0.06656716018915176, + 0.029578251764178276, + 0.028423432260751724, + 0.024671347811818123, + 0.052931420505046844, + 0.008854569867253304, + 0.014263145625591278, + 0.007130863144993782, + 0.016288956627249718, + 0.005861947312951088, + 0.008160176686942577, + 0.006768906954675913, + 0.01714155077934265, + 0.00731880497187376, + 0.005820535589009523, + 0.005551824811846018, + 0.015874730423092842, + 0.010723575949668884, + 0.008739065378904343, + 0.008803553879261017, + 0.01231548935174942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.00675099017098546 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0860178992152214, + 0.0775555819272995, + 0.050858862698078156, + 0.05592113360762596, + 0.09899512678384781, + 0.08055621385574341, + 0.25090038776397705, + 0.18877458572387695, + 0.041964251548051834, + 0.06845588982105255 + ], + "credit_map": [ + 0.009599956683814526, + 0.00971358921378851, + 0.010894110426306725, + 0.015393505804240704, + 0.010375081561505795, + 0.013450238853693008, + 0.007901913486421108, + 0.008689513429999352, + 0.011914916336536407, + 0.007381086703389883, + 0.007235141005367041, + 0.012994352728128433, + 0.009782465174794197, + 0.006647540722042322, + 0.012351096607744694, + 0.009248990565538406, + 0.00590460654348135, + 0.004848182667046785, + 0.006920608226209879, + 0.005689443554729223, + 0.00460712518543005, + 0.006734916009008884, + 0.010430846363306046, + 0.005723142065107822, + 0.0056860498152673244, + 0.009290465153753757, + 0.008323900401592255, + 0.008133932948112488, + 0.007981893606483936, + 0.0035000217612832785, + 0.007043217774480581, + 0.005961655173450708, + 0.011058823205530643, + 0.012058595195412636, + 0.015586553141474724, + 0.014507547952234745, + 0.010873911902308464, + 0.009873089380562305, + 0.01622198522090912, + 0.008814625442028046, + 0.008272942155599594, + 0.01190177258104086, + 0.007311577908694744, + 0.01256321370601654, + 0.005212509073317051, + 0.01392432115972042, + 0.009926038794219494, + 0.011443844996392727, + 0.018201544880867004, + 0.024125507101416588, + 0.022267330437898636, + 0.04674699157476425, + 0.031531963497400284, + 0.03033881075680256, + 0.038255952298641205, + 0.03943230211734772, + 0.031021086499094963, + 0.02759370394051075, + 0.02811657451093197, + 0.028941046446561813, + 0.015012961812317371, + 0.015007265843451023, + 0.01540420949459076, + 0.027677757665514946, + 0.0036839491222053766, + 0.00714844698086381, + 0.0031713824719190598, + 0.006642598658800125, + 0.003644135082140565, + 0.006471806205809116, + 0.004775893408805132, + 0.006426042877137661, + 0.004987149965018034, + 0.009288672357797623, + 0.00541874160990119, + 0.01219722256064415, + 0.007274075411260128, + 0.009763692505657673, + 0.005752935074269772, + 0.013773409649729729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.00978058110922575 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705665826797485 + ], + "opposite_logprob_tokens": [ + -2.12705659866333 + ], + "logprob_reward": -0.12705665826797485, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705665826797485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09178391844034195, + 0.05448223650455475, + 0.05092325061559677, + 0.07893216609954834, + 0.1364756077528, + 0.19382977485656738, + 0.06476067006587982, + 0.13612768054008484, + 0.11922718584537506, + 0.07345745712518692 + ], + "credit_map": [ + 0.015111120417714119, + 0.0128913139924407, + 0.01449936255812645, + 0.010739046148955822, + 0.01011717040091753, + 0.009139838628470898, + 0.00948342029005289, + 0.009802643209695816, + 0.00519063463434577, + 0.004661219194531441, + 0.006876541301608086, + 0.007879430428147316, + 0.005835165735334158, + 0.005831929389387369, + 0.011494581587612629, + 0.006712733767926693, + 0.004245294723659754, + 0.003391612321138382, + 0.006204782985150814, + 0.0056481873616576195, + 0.005115445237606764, + 0.006821206770837307, + 0.008421429432928562, + 0.011075292713940144, + 0.006482299882918596, + 0.01241660863161087, + 0.007096904795616865, + 0.009030042216181755, + 0.014319680631160736, + 0.006268802098929882, + 0.011056704446673393, + 0.012261123396456242, + 0.014174971729516983, + 0.014909736812114716, + 0.016972286626696587, + 0.021638497710227966, + 0.014375943690538406, + 0.015720704570412636, + 0.023575469851493835, + 0.015108001418411732, + 0.018926387652754784, + 0.018521437421441078, + 0.017308345064520836, + 0.03846448287367821, + 0.013232776895165443, + 0.022420810535550117, + 0.019721103832125664, + 0.04523444175720215, + 0.0120155094191432, + 0.004911580588668585, + 0.012810274958610535, + 0.006634725723415613, + 0.005876769777387381, + 0.00807193759828806, + 0.004856100305914879, + 0.009583772160112858, + 0.012312546372413635, + 0.010211802087724209, + 0.010885915718972683, + 0.03384579345583916, + 0.009717671200633049, + 0.0118940444663167, + 0.007723786402493715, + 0.03953612223267555, + 0.01209256611764431, + 0.018901992589235306, + 0.010897226631641388, + 0.022845597937703133, + 0.011326203122735023, + 0.010740064084529877, + 0.011915159411728382, + 0.020508382469415665, + 0.0070891487412154675, + 0.011222433298826218, + 0.00658420380204916, + 0.012760923244059086, + 0.008388364687561989, + 0.008481093682348728, + 0.005459261126816273, + 0.013472029939293861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.12705665826797485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, subjective point-of-view shot, shot from inside a car, reveals a side window covered in dense raindrops. Each droplet acts like a miniature lens, distorting and refracting the world outside. The raindrops slowly slide down the glass, leaving behind blurred streaks that add a hazy, tense quality to the scene. Outside, a wet city alley stretches out into the night. The asphalt pavement reflects the flickering red, blue, and green neon lights of the surrounding shop signs, creating splatters of oscillating watercolor. Amidst this dazzling yet chilling light, two men stand face to face: one is tall, clad in a long, dark trench coat, its soaked fabric clinging to his tense shoulders. He leans slightly forward, his posture imbued with oppression, his facial expression sullen yet resolute in the shifting light. The other, younger, wears a black hoodie, his face half hidden in shadow, half illuminated by a blue neon glow. His arms are slightly outstretched, his hands clenched into half fists, his eyes a mixture of vigilance and unyielding determination. They stand a few steps apart, motionless, yet the air is thick with tension, as if a confrontation could erupt at any moment. The entire scene, shrouded in cold tones and high-contrast lighting, creates a cinematic experience of profound suspense and uncertainty.", + "seed": 3709, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_6/index_4636_seed_3709_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": -0.033048281259834764, + "vlm_logprob": -0.033048281259834764, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.033048281259834764, + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029275409877300262 + ], + "opposite_logprob_tokens": [ + -3.904275417327881 + ], + "logprob_reward": -0.029275409877300262, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029275409877300262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18650278449058533, + 0.20912955701351166, + 0.04250066727399826, + 0.1042526513338089, + 0.06435303390026093, + 0.056794919073581696, + 0.04357995465397835, + 0.026608014479279518, + 0.20602209866046906, + 0.06025633215904236 + ], + "credit_map": [ + 0.020955713465809822, + 0.01931336335837841, + 0.017171885818243027, + 0.034225549548864365, + 0.016549933701753616, + 0.025545258074998856, + 0.026648100465536118, + 0.026092970743775368, + 0.02108456939458847, + 0.020348137244582176, + 0.03679032623767853, + 0.028968557715415955, + 0.014509879983961582, + 0.032401494681835175, + 0.023017987608909607, + 0.03200862184166908, + 0.004779206123203039, + 0.0033314295578747988, + 0.006596601102501154, + 0.004590146709233522, + 0.0033053902443498373, + 0.010652797296643257, + 0.004487966187298298, + 0.004757129121571779, + 0.014689106494188309, + 0.016070840880274773, + 0.010511552914977074, + 0.019962575286626816, + 0.005538781173527241, + 0.008027958683669567, + 0.004769069608300924, + 0.02468276396393776, + 0.00514253182336688, + 0.0072776745073497295, + 0.006045930087566376, + 0.0118933180347085, + 0.0061132945120334625, + 0.0073638311587274075, + 0.008111144416034222, + 0.012405305169522762, + 0.008639913983643055, + 0.007641006726771593, + 0.005565366242080927, + 0.00866756122559309, + 0.006465619429945946, + 0.00811244361102581, + 0.005237926263362169, + 0.006465086247771978, + 0.004633746109902859, + 0.0072393943555653095, + 0.004969772417098284, + 0.007942000404000282, + 0.002830810146406293, + 0.004122635815292597, + 0.0028920667245984077, + 0.008949529379606247, + 0.002871305448934436, + 0.0043883551843464375, + 0.0026374852750450373, + 0.0051018353551626205, + 0.0017822005320340395, + 0.0024466172326356173, + 0.0015174980508163571, + 0.005862717982381582, + 0.02939382940530777, + 0.012128673493862152, + 0.0424245148897171, + 0.033058587461709976, + 0.008182100020349026, + 0.01478577870875597, + 0.012496679089963436, + 0.05355193838477135, + 0.005942931864410639, + 0.006167140323668718, + 0.009568874724209309, + 0.009555837139487267, + 0.006692965514957905, + 0.006586547940969467, + 0.0037905611097812653, + 0.011951472610235214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.029275409877300262 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020615356042981148 + ], + "opposite_logprob_tokens": [ + -3.895615339279175 + ], + "logprob_reward": -0.020615356042981148, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020615356042981148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1092258095741272, + 0.05443042144179344, + 0.0222089271992445, + 0.1078513115644455, + 0.09460359811782837, + 0.10080644488334656, + 0.09883750975131989, + 0.06608448177576065, + 0.28168538212776184, + 0.06426607072353363 + ], + "credit_map": [ + 0.010749324224889278, + 0.013377386145293713, + 0.011558018624782562, + 0.018959257751703262, + 0.007673601154237986, + 0.012905526906251907, + 0.015350403264164925, + 0.018652301281690598, + 0.004759976640343666, + 0.005651269108057022, + 0.005726470611989498, + 0.007724070455878973, + 0.004983160179108381, + 0.009426559321582317, + 0.007241467013955116, + 0.008917446248233318, + 0.00322828465141356, + 0.0020498884841799736, + 0.002211768412962556, + 0.0032431764993816614, + 0.0016029936959967017, + 0.0051378714852035046, + 0.002175749745219946, + 0.0025591931771486998, + 0.015263649635016918, + 0.012389728799462318, + 0.015163526870310307, + 0.0197344571352005, + 0.008515030145645142, + 0.010104851797223091, + 0.0073733702301979065, + 0.01930670440196991, + 0.008597513660788536, + 0.014482193626463413, + 0.013861147686839104, + 0.01607709378004074, + 0.00815517921000719, + 0.009272574447095394, + 0.009139565750956535, + 0.01501833088696003, + 0.012793955393135548, + 0.012754308059811592, + 0.013389096595346928, + 0.017150552943348885, + 0.009916716255247593, + 0.015427905134856701, + 0.006995271425694227, + 0.012378634884953499, + 0.008996984921395779, + 0.019863078370690346, + 0.011283366940915585, + 0.021075112745165825, + 0.006356546655297279, + 0.009091941639780998, + 0.006375561002641916, + 0.015794917941093445, + 0.007662043906748295, + 0.00792420469224453, + 0.00556976580992341, + 0.01288648322224617, + 0.006376035511493683, + 0.008533944375813007, + 0.005318041425198317, + 0.011813957244157791, + 0.027356483042240143, + 0.009930609725415707, + 0.0684858039021492, + 0.04967525228857994, + 0.006422100588679314, + 0.017468683421611786, + 0.013525634072721004, + 0.08882081508636475, + 0.006944811437278986, + 0.0069605158641934395, + 0.012234112247824669, + 0.009413095191121101, + 0.004502107854932547, + 0.008067304268479347, + 0.0047197844833135605, + 0.011424335651099682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.020615356042981148 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259605124592781 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.005259605124592781, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259605124592781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04472065716981888, + 0.02013634890317917, + 0.006421157158911228, + 0.012444581836462021, + 0.019125839695334435, + 0.052012402564287186, + 0.04723722115159035, + 0.023635024204850197, + 0.7391640543937683, + 0.035102665424346924 + ], + "credit_map": [ + 0.004077360965311527, + 0.005624052602797747, + 0.003409395692870021, + 0.008262737654149532, + 0.004632123280316591, + 0.004723664838820696, + 0.00679765222594142, + 0.0071936738677322865, + 0.002344524022191763, + 0.0015809445176273584, + 0.005112845916301012, + 0.0028116696048527956, + 0.001692660036496818, + 0.002115745795890689, + 0.0016256604576483369, + 0.002852298552170396, + 0.0011032411130145192, + 0.0006724704871885478, + 0.0010177503572776914, + 0.0009216759353876114, + 0.0004909982671961188, + 0.0007366097415797412, + 0.000612416013609618, + 0.0008659959421493113, + 0.0013460242189466953, + 0.0011873828480020165, + 0.0029812855646014214, + 0.0017540110275149345, + 0.0008553387597203255, + 0.0014840211952105165, + 0.000939668738283217, + 0.0018968507647514343, + 0.0015711872838437557, + 0.002988447668030858, + 0.002338611287996173, + 0.0038163012359291315, + 0.0017724520293995738, + 0.002017238875851035, + 0.002493897220119834, + 0.0021277060732245445, + 0.005171485710889101, + 0.008735881187021732, + 0.005774805322289467, + 0.01028564851731062, + 0.004313354380428791, + 0.00732179032638669, + 0.0033667476382106543, + 0.0070426929742097855, + 0.004277834203094244, + 0.008839893154799938, + 0.005172389559447765, + 0.010160931386053562, + 0.003483427222818136, + 0.004251513164490461, + 0.0031561071518808603, + 0.007895125076174736, + 0.0018889149650931358, + 0.0034996229223906994, + 0.0020668855868279934, + 0.005156440194696188, + 0.0014100174885243177, + 0.0017052246257662773, + 0.0013397707371041179, + 0.006568149663507938, + 0.07185360789299011, + 0.020466983318328857, + 0.18929660320281982, + 0.13346847891807556, + 0.012262082658708096, + 0.03881356492638588, + 0.03253239765763283, + 0.24047036468982697, + 0.003864285070449114, + 0.0035565809812396765, + 0.006480658892542124, + 0.006034423131495714, + 0.002150703687220812, + 0.003230307251214981, + 0.0014846025733277202, + 0.008301107212901115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.005259605124592781 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977963674813509 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.00977963674813509, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977963674813509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04489176720380783, + 0.0261701550334692, + 0.007847306318581104, + 0.02799353003501892, + 0.02829914167523384, + 0.04236726835370064, + 0.012209595181047916, + 0.015220828354358673, + 0.7855972051620483, + 0.009403139352798462 + ], + "credit_map": [ + 0.0048818280920386314, + 0.005713714752346277, + 0.004835410043597221, + 0.008431585505604744, + 0.004576234146952629, + 0.005806147586554289, + 0.005361270159482956, + 0.005285575520247221, + 0.003188959090039134, + 0.0034722324926406145, + 0.00461700651794672, + 0.00374806672334671, + 0.002485969103872776, + 0.002189067890867591, + 0.0027782348915934563, + 0.0036906194873154163, + 0.0012262659147381783, + 0.0009462455636821687, + 0.0011321644997224212, + 0.00109785795211792, + 0.0006807994213886559, + 0.0008071560878306627, + 0.0008515283698216081, + 0.0011052886256948113, + 0.0031138078775256872, + 0.004598624538630247, + 0.004312078934162855, + 0.004523457959294319, + 0.002803252311423421, + 0.002701176330447197, + 0.0025154189206659794, + 0.003425714559853077, + 0.00201981607824564, + 0.004353615455329418, + 0.0037013927940279245, + 0.005778958089649677, + 0.002403072314336896, + 0.003390373894944787, + 0.002608692739158869, + 0.0040432182140648365, + 0.006198045797646046, + 0.0046107154339551926, + 0.005830795969814062, + 0.006983119994401932, + 0.003757716855034232, + 0.004829760175198317, + 0.004056286066770554, + 0.006100830156356096, + 0.0015674229944124818, + 0.001929286285303533, + 0.001095342100597918, + 0.002061875071376562, + 0.0009431821526959538, + 0.0019471575506031513, + 0.0010306391632184386, + 0.0016346898628398776, + 0.0014531207270920277, + 0.0029661529697477818, + 0.0010348265059292316, + 0.0032336730509996414, + 0.0012718306388705969, + 0.0014973494689911604, + 0.0008902260451577604, + 0.002873648889362812, + 0.07577156275510788, + 0.021557282656431198, + 0.2011445164680481, + 0.1420493721961975, + 0.012953587807714939, + 0.04124707728624344, + 0.034535691142082214, + 0.2563381791114807, + 0.000820422894321382, + 0.0012582010822370648, + 0.0010252778884023428, + 0.002228621393442154, + 0.0006380951381288469, + 0.0011659171432256699, + 0.0007726690964773297, + 0.001493934658356011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.00977963674813509 + }, + { + "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 from inside a car looking out a side window covered in dense raindrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two men standing face to face outside in a wet city alley at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one man wearing a long dark trench coat and leaning slightly forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the other man wearing a black hoodie with his face half hidden in shadow and half illuminated by blue neon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the characters remain motionless while the raindrops slide down the glass and neon lights reflect on the wet pavement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031139850616455 + ], + "opposite_logprob_tokens": [ + -2.350311279296875 + ], + "logprob_reward": -0.10031139850616455, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031139850616455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025181317701935768, + 0.010287128388881683, + 0.0047677913680672646, + 0.00910450704395771, + 0.012036382220685482, + 0.017176618799567223, + 0.007102452218532562, + 0.0067746080458164215, + 0.8975512385368347, + 0.01001792773604393 + ], + "credit_map": [ + 0.00252900761552155, + 0.002111785812303424, + 0.0029259154107421637, + 0.004106801003217697, + 0.0019575359765440226, + 0.0037528187967836857, + 0.0032661347649991512, + 0.004531318787485361, + 0.0010102156084030867, + 0.0007838357705622911, + 0.0015647296095266938, + 0.0015982449986040592, + 0.0009085359633900225, + 0.0015364993596449494, + 0.0012671693693846464, + 0.0016178976511582732, + 0.0004928107955493033, + 0.0004043796507176012, + 0.0012768596643581986, + 0.0005909641040489078, + 0.00035153882345184684, + 0.0007117401692084968, + 0.0003828611224889755, + 0.0005566370091401041, + 0.0013911717105656862, + 0.0016316461842507124, + 0.0008534085936844349, + 0.001460328116081655, + 0.0007835233118385077, + 0.000991820590570569, + 0.0006641336367465556, + 0.0013284749584272504, + 0.0011931689223274589, + 0.001629440812394023, + 0.002185711171478033, + 0.002069998299703002, + 0.0008954128716140985, + 0.0016460625920444727, + 0.0009271154995076358, + 0.0014894724590703845, + 0.002200718969106674, + 0.0024680616334080696, + 0.001827905187383294, + 0.0031687873415648937, + 0.0014672301476821303, + 0.002810250036418438, + 0.0012128176167607307, + 0.002020846353843808, + 0.0007393679697997868, + 0.0012602832866832614, + 0.0005561769939959049, + 0.0014688128139823675, + 0.00047931328299455345, + 0.0007443818612955511, + 0.0007915713940747082, + 0.0010625452268868685, + 0.0007504009990952909, + 0.0007587869768030941, + 0.0007522343075834215, + 0.0013579238438978791, + 0.00039345407276414335, + 0.000734075321815908, + 0.0006304650451056659, + 0.0013972674496471882, + 0.08753491938114166, + 0.02486630529165268, + 0.2296564131975174, + 0.16180448234081268, + 0.01468953862786293, + 0.04740909859538078, + 0.03962887078523636, + 0.2919616401195526, + 0.0009605558589100838, + 0.0009864156600087881, + 0.0016607630532234907, + 0.001960424706339836, + 0.0008837670902721584, + 0.0011818541679531336, + 0.0006568924873135984, + 0.0017272551776841283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.10031139850616455 + } + ], + "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/0000167/group_7/index_tensor([3069, 3069])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_tensor([3069, 3069])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e73c0ebdbb528298b1abced9c9b81617ca766799 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_tensor([3069, 3069])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a powerful, overhead close-up, the camera slowly zooms in, closing in on the behemoth Godzilla. Under a pitch-black night sky, shrouded in lightning and thunder, it stands amid the rubble of a city, surrounded by twisted steel and shattered concrete. The air is thick with smoke and ash. Godzilla's armor, clad in deep black scales as rough and battle-worn as ancient rock, stares forward with eyes blazing with atomic blue fury. A maw studded with dagger-sharp teeth suddenly opens, and a deafening roar erupts from its deep throat. Its massive claws loom menacingly at the edge of the frame. The force of the sonic boom distorts the air, sending rubble from the crumbling buildings around it careening down. At that moment, a sharp blue anamorphic lens flare slowly and gracefully streaks across its scarred, hideous face, briefly illuminating the rough texture of its skin and the utter fury of its expression. The entire scene is illuminated by distant city flames and occasional flashes of lightning, creating an apocalyptic, tense atmosphere filled with destructive power and primal rage.", + "seed": 10994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10994_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": -0.11265861913561821, + "vlm_logprob": -0.11265861913561821, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.11265861913561821, + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3051355481147766 + ], + "opposite_logprob_tokens": [ + -1.4301354885101318 + ], + "logprob_reward": -0.3051355481147766, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3051355481147766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0974016785621643, + 0.19580808281898499, + 0.111601322889328, + 0.23585988581180573, + 0.10069158673286438, + 0.0795852392911911, + 0.0480387844145298, + 0.033415745943784714, + 0.04756812006235123, + 0.05002957955002785 + ], + "credit_map": [ + 0.011256298050284386, + 0.012443493120372295, + 0.01232474111020565, + 0.006317725405097008, + 0.01627454347908497, + 0.009196985512971878, + 0.02147754468023777, + 0.008110354654490948, + 0.018141865730285645, + 0.02345586009323597, + 0.040047165006399155, + 0.01696242019534111, + 0.022649714723229408, + 0.021116113290190697, + 0.039569348096847534, + 0.01386559009552002, + 0.02221187949180603, + 0.015006566420197487, + 0.015218875370919704, + 0.009727127850055695, + 0.010210041888058186, + 0.014330658130347729, + 0.011100736446678638, + 0.013795447535812855, + 0.05914970859885216, + 0.01872597448527813, + 0.052967704832553864, + 0.020009944215416908, + 0.01163843460381031, + 0.034075651317834854, + 0.02061746083199978, + 0.018674982711672783, + 0.01520528458058834, + 0.02236637845635414, + 0.01571623980998993, + 0.014875810593366623, + 0.007607278414070606, + 0.0066544534638524055, + 0.006907808594405651, + 0.011358330957591534, + 0.008822840638458729, + 0.00814212579280138, + 0.0065118358470499516, + 0.011512177996337414, + 0.008199612610042095, + 0.014558295719325542, + 0.01102080475538969, + 0.010817543603479862, + 0.007347461301833391, + 0.00611236784607172, + 0.006639204919338226, + 0.00563813978806138, + 0.004219484515488148, + 0.0074588842689991, + 0.005389641970396042, + 0.0052335988730192184, + 0.005097452085465193, + 0.005772685632109642, + 0.0032561065163463354, + 0.00537910358980298, + 0.0028895435389131308, + 0.003764219582080841, + 0.0027445536106824875, + 0.004512083251029253, + 0.0064617968164384365, + 0.0038958252407610416, + 0.005881837103515863, + 0.007665068376809359, + 0.004579076077789068, + 0.004603833891451359, + 0.006948086433112621, + 0.007532595191150904, + 0.00725353229790926, + 0.006028750911355019, + 0.006378470454365015, + 0.005835307762026787, + 0.007322204764932394, + 0.0031125866807997227, + 0.009527179412543774, + 0.004571547731757164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.3051355481147766 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991693317890167 + ], + "opposite_logprob_tokens": [ + -1.8049169778823853 + ], + "logprob_reward": -0.17991693317890167, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991693317890167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07390301674604416, + 0.22638489305973053, + 0.1723281890153885, + 0.12035716325044632, + 0.11520681530237198, + 0.07482679188251495, + 0.10546542704105377, + 0.07933945953845978, + 0.01911516860127449, + 0.013073064386844635 + ], + "credit_map": [ + 0.008718454279005527, + 0.009128205478191376, + 0.011096588335931301, + 0.00885863322764635, + 0.009645994752645493, + 0.0065560368821024895, + 0.013254568912088871, + 0.0066445390693843365, + 0.034957289695739746, + 0.033779360353946686, + 0.026167944073677063, + 0.03172675520181656, + 0.029466405510902405, + 0.027146628126502037, + 0.024715125560760498, + 0.01842537149786949, + 0.054281480610370636, + 0.015123731456696987, + 0.008786912076175213, + 0.008141872473061085, + 0.0071245161816477776, + 0.04754801094532013, + 0.007998534478247166, + 0.02332313545048237, + 0.019621871411800385, + 0.01594538614153862, + 0.018339797854423523, + 0.016819879412651062, + 0.009479719214141369, + 0.011487766169011593, + 0.01610514707863331, + 0.012557589448988438, + 0.014876858331263065, + 0.02509739063680172, + 0.016757573932409286, + 0.014794422313570976, + 0.009283277206122875, + 0.008429735898971558, + 0.008677460253238678, + 0.01729009486734867, + 0.007819021120667458, + 0.007274126634001732, + 0.006850419566035271, + 0.010038730688393116, + 0.005994916893541813, + 0.01574929617345333, + 0.009763690643012524, + 0.011336589232087135, + 0.014739885926246643, + 0.016851574182510376, + 0.010182059369981289, + 0.016993537545204163, + 0.009473293088376522, + 0.012958220206201077, + 0.010875179432332516, + 0.013391679152846336, + 0.012573057785630226, + 0.012328019365668297, + 0.007498129270970821, + 0.01224159449338913, + 0.006781071424484253, + 0.008742856793105602, + 0.008501117117702961, + 0.010673610493540764, + 0.0034551729913800955, + 0.0019438706804066896, + 0.002285382943227887, + 0.0026534132193773985, + 0.0020077049266546965, + 0.0013987879501655698, + 0.002234499901533127, + 0.0031363354064524174, + 0.0017455726629123092, + 0.0014069615863263607, + 0.001774517702870071, + 0.0015684581594541669, + 0.0015731357270851731, + 0.0015116888098418713, + 0.002086877590045333, + 0.0014058516826480627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.17991693317890167 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196435645222664 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196435645222664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196435645222664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716869562864304, + 0.19078880548477173, + 0.08260481059551239, + 0.1173677146434784, + 0.08142059296369553, + 0.18432946503162384, + 0.09870518743991852, + 0.08392884582281113, + 0.04204927757382393, + 0.041636593639850616 + ], + "credit_map": [ + 0.010314487852156162, + 0.013740292750298977, + 0.009810587391257286, + 0.006918470840901136, + 0.013763424009084702, + 0.007006046362221241, + 0.010615499690175056, + 0.004999895580112934, + 0.029638495296239853, + 0.023805513978004456, + 0.021708769723773003, + 0.02422068454325199, + 0.02277311310172081, + 0.024803509935736656, + 0.02385035529732704, + 0.019988372921943665, + 0.013638763688504696, + 0.00907105766236782, + 0.009066316299140453, + 0.010749788023531437, + 0.006611563265323639, + 0.01369024720042944, + 0.00837843120098114, + 0.011398646980524063, + 0.026582011952996254, + 0.012393456883728504, + 0.02412792108952999, + 0.013536208309233189, + 0.008866160176694393, + 0.01107992883771658, + 0.014029784128069878, + 0.006752250716090202, + 0.010224828496575356, + 0.018685052171349525, + 0.013155645690858364, + 0.01054464466869831, + 0.007623151410371065, + 0.006270600948482752, + 0.0059925466775894165, + 0.00892412569373846, + 0.02294294349849224, + 0.018124500289559364, + 0.02896374836564064, + 0.02448742650449276, + 0.01443121861666441, + 0.028724532574415207, + 0.021984979510307312, + 0.024670127779245377, + 0.0105205113068223, + 0.016427326947450638, + 0.015585403889417648, + 0.01294031459838152, + 0.0064773401245474815, + 0.011333229951560497, + 0.012012098915874958, + 0.013408975675702095, + 0.011731432750821114, + 0.016046227887272835, + 0.007064524572342634, + 0.01234931219369173, + 0.007519020698964596, + 0.00849502719938755, + 0.008574866689741611, + 0.012148444540798664, + 0.006213176995515823, + 0.004189397674053907, + 0.007455709856003523, + 0.0058168405666947365, + 0.0035331547260284424, + 0.003247890854254365, + 0.004940141458064318, + 0.0066529675386846066, + 0.0057050264440476894, + 0.005376045126467943, + 0.00569880148395896, + 0.004950417205691338, + 0.005078102461993694, + 0.00399131840094924, + 0.006622401997447014, + 0.004214482381939888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.018196435645222664 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07228638231754303, + 0.08598426729440689, + 0.11421164870262146, + 0.13243572413921356, + 0.09962733834981918, + 0.26689988374710083, + 0.12725353240966797, + 0.04409841075539589, + 0.03200938552618027, + 0.02519341930747032 + ], + "credit_map": [ + 0.008625970222055912, + 0.009983442723751068, + 0.014399235136806965, + 0.008196191862225533, + 0.009743739850819111, + 0.005729702766984701, + 0.009413058869540691, + 0.006195043679326773, + 0.009797599166631699, + 0.010251778177917004, + 0.013719575479626656, + 0.010419060476124287, + 0.008776962757110596, + 0.014598168432712555, + 0.013636437244713306, + 0.004784683231264353, + 0.03262580186128616, + 0.011498833075165749, + 0.010620791465044022, + 0.006159418728202581, + 0.010028279386460781, + 0.02309916540980339, + 0.007640847470611334, + 0.012538515962660313, + 0.020608916878700256, + 0.011344769969582558, + 0.024480128660798073, + 0.01657266542315483, + 0.010026969015598297, + 0.015283110551536083, + 0.01979116164147854, + 0.014328010380268097, + 0.012559235095977783, + 0.025531291961669922, + 0.01732179895043373, + 0.011379712261259556, + 0.008231068961322308, + 0.0074851918034255505, + 0.00699954479932785, + 0.010119494050741196, + 0.02665914222598076, + 0.023664144799113274, + 0.03680562600493431, + 0.03852563723921776, + 0.018557587638497353, + 0.05590273067355156, + 0.030411990359425545, + 0.03637302294373512, + 0.016156353056430817, + 0.014753601513803005, + 0.013970502652227879, + 0.016071699559688568, + 0.016167446970939636, + 0.016463628038764, + 0.020647583529353142, + 0.013022707775235176, + 0.006899493280798197, + 0.005030023865401745, + 0.004080954007804394, + 0.006023728288710117, + 0.004656580276787281, + 0.0036541931331157684, + 0.0072482880204916, + 0.006505150347948074, + 0.004184059798717499, + 0.004245958290994167, + 0.005833981558680534, + 0.003746027359738946, + 0.0031475627329200506, + 0.002602730644866824, + 0.003676456166431308, + 0.004572609905153513, + 0.00386481243185699, + 0.00312424311414361, + 0.003425794653594494, + 0.0017864578403532505, + 0.003766490612179041, + 0.0026652575470507145, + 0.004107515327632427, + 0.0024528487119823694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.03369844704866409 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634573169052601 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634573169052601, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634573169052601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0588986799120903, + 0.10973818600177765, + 0.1194717213511467, + 0.1285204142332077, + 0.11731570214033127, + 0.1890849620103836, + 0.10558690875768661, + 0.05708491802215576, + 0.06419927626848221, + 0.0500992052257061 + ], + "credit_map": [ + 0.008685259148478508, + 0.008144522085785866, + 0.010198209434747696, + 0.0058588129468262196, + 0.0067250896245241165, + 0.004677849356085062, + 0.008897055871784687, + 0.005711881909519434, + 0.01512082014232874, + 0.012533606961369514, + 0.016135558485984802, + 0.014798346906900406, + 0.010932104662060738, + 0.018609369173645973, + 0.014734596945345402, + 0.0068737817928195, + 0.03464179486036301, + 0.011570150032639503, + 0.00906192883849144, + 0.006787810008972883, + 0.0072696758434176445, + 0.02935604192316532, + 0.006251614540815353, + 0.014532697387039661, + 0.016389688476920128, + 0.02002982422709465, + 0.017704904079437256, + 0.016803327947854996, + 0.012071874924004078, + 0.011876307427883148, + 0.021469425410032272, + 0.012175058014690876, + 0.01494570542126894, + 0.027251847088336945, + 0.019843874499201775, + 0.01353871263563633, + 0.009665980003774166, + 0.008622225373983383, + 0.0076971473172307014, + 0.015750207006931305, + 0.01956324279308319, + 0.016728868708014488, + 0.028284166008234024, + 0.025789476931095123, + 0.013500869274139404, + 0.036439958959817886, + 0.021168243139982224, + 0.027610139921307564, + 0.012657590210437775, + 0.015599156729876995, + 0.00807531364262104, + 0.015363892540335655, + 0.011633947491645813, + 0.011568807996809483, + 0.018871793523430824, + 0.011816403828561306, + 0.010477838106453419, + 0.006022259593009949, + 0.006216933019459248, + 0.008259099908173084, + 0.005100637208670378, + 0.00447356840595603, + 0.009284044615924358, + 0.007250538561493158, + 0.011143450625240803, + 0.004569091834127903, + 0.008523807860910892, + 0.008047978393733501, + 0.008418897166848183, + 0.005884776823222637, + 0.007252472918480635, + 0.010358799248933792, + 0.006743432022631168, + 0.006275842431932688, + 0.006983734667301178, + 0.005957921966910362, + 0.006822575815021992, + 0.004757575690746307, + 0.007612899411469698, + 0.004945223685353994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.02634573169052601 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, overhead close-up, the camera slowly zooms in, closing in on the behemoth Godzilla. Under a pitch-black night sky, shrouded in lightning and thunder, it stands amid the rubble of a city, surrounded by twisted steel and shattered concrete. The air is thick with smoke and ash. Godzilla's armor, clad in deep black scales as rough and battle-worn as ancient rock, stares forward with eyes blazing with atomic blue fury. A maw studded with dagger-sharp teeth suddenly opens, and a deafening roar erupts from its deep throat. Its massive claws loom menacingly at the edge of the frame. The force of the sonic boom distorts the air, sending rubble from the crumbling buildings around it careening down. At that moment, a sharp blue anamorphic lens flare slowly and gracefully streaks across its scarred, hideous face, briefly illuminating the rough texture of its skin and the utter fury of its expression. The entire scene is illuminated by distant city flames and occasional flashes of lightning, creating an apocalyptic, tense atmosphere filled with destructive power and primal rage.", + "seed": 10995, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10995_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": -0.15894427858293056, + "vlm_logprob": -0.15894427858293056, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.15894427858293056, + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.712050199508667 + ], + "opposite_logprob_tokens": [ + -0.712050199508667 + ], + "logprob_reward": -0.712050199508667, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.712050199508667, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13557551801204681, + 0.10314520448446274, + 0.15465664863586426, + 0.11244949698448181, + 0.06785115599632263, + 0.05616094172000885, + 0.10893157869577408, + 0.14049620926380157, + 0.04794510826468468, + 0.07278813421726227 + ], + "credit_map": [ + 0.014403359964489937, + 0.019281374290585518, + 0.014572742395102978, + 0.02100084163248539, + 0.022322695702314377, + 0.018032517284154892, + 0.015059011057019234, + 0.01090298406779766, + 0.008901245892047882, + 0.012464909814298153, + 0.005458000581711531, + 0.004686674568802118, + 0.029763581231236458, + 0.02954954095184803, + 0.006206674501299858, + 0.00611457135528326, + 0.011841715313494205, + 0.01646481081843376, + 0.02047797292470932, + 0.025107376277446747, + 0.018142487853765488, + 0.014584452845156193, + 0.028713122010231018, + 0.019324710592627525, + 0.010622060857713223, + 0.02398304082453251, + 0.00870121642947197, + 0.01113172061741352, + 0.015601197257637978, + 0.016720913350582123, + 0.015132775530219078, + 0.01055657397955656, + 0.004336802754551172, + 0.005418058484792709, + 0.0033724328968673944, + 0.020566770806908607, + 0.004403308499604464, + 0.02437054179608822, + 0.0029820792842656374, + 0.0024011635687202215, + 0.00859226007014513, + 0.009032927453517914, + 0.0049901423044502735, + 0.010599072091281414, + 0.005814732983708382, + 0.006515537854284048, + 0.005552208982408047, + 0.005064061842858791, + 0.023137537762522697, + 0.010827729478478432, + 0.009098689071834087, + 0.011170848272740841, + 0.011725613847374916, + 0.013659332878887653, + 0.01849900372326374, + 0.010812819935381413, + 0.01767968200147152, + 0.02063603512942791, + 0.013086790218949318, + 0.02211807109415531, + 0.01918003521859646, + 0.017092060297727585, + 0.013715038076043129, + 0.016988489776849747, + 0.0067058163695037365, + 0.005885718390345573, + 0.0036312805023044348, + 0.005679604131728411, + 0.0077199977822601795, + 0.006919474340975285, + 0.00604022853076458, + 0.005362987518310547, + 0.009231003932654858, + 0.005596654489636421, + 0.013763293623924255, + 0.007863162085413933, + 0.0053763664327561855, + 0.011261395178735256, + 0.00999117735773325, + 0.009705078788101673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.712050199508667 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020613020285964012 + ], + "opposite_logprob_tokens": [ + -3.895612955093384 + ], + "logprob_reward": -0.020613020285964012, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020613020285964012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05911044031381607, + 0.07565465569496155, + 0.13481785356998444, + 0.033136025071144104, + 0.14610284566879272, + 0.0374373160302639, + 0.04410420358181, + 0.3478281795978546, + 0.058413151651620865, + 0.06339537352323532 + ], + "credit_map": [ + 0.005588581785559654, + 0.012941541150212288, + 0.005733550991863012, + 0.010578623041510582, + 0.006232674699276686, + 0.007707685697823763, + 0.006155035458505154, + 0.004172749351710081, + 0.007206387352198362, + 0.016874326393008232, + 0.006122787948697805, + 0.005559029523283243, + 0.010198233649134636, + 0.006706681102514267, + 0.013162356801331043, + 0.009824855253100395, + 0.011743098497390747, + 0.012148662470281124, + 0.020331423729658127, + 0.023618798702955246, + 0.01430516131222248, + 0.008230564184486866, + 0.028531285002827644, + 0.015908855944871902, + 0.0023826684337109327, + 0.007720078341662884, + 0.0026272349059581757, + 0.005219753831624985, + 0.0031133294105529785, + 0.003666104981675744, + 0.005607262719422579, + 0.0027995912823826075, + 0.009470304474234581, + 0.010296155698597431, + 0.006253750529140234, + 0.041613709181547165, + 0.010398413054645061, + 0.0532451756298542, + 0.0059680393896996975, + 0.008857294917106628, + 0.006431848742067814, + 0.0061013177037239075, + 0.0033484187442809343, + 0.00533869955688715, + 0.0026328808162361383, + 0.0039239381439983845, + 0.005623833741992712, + 0.004036378115415573, + 0.0062791150994598866, + 0.0048781163059175014, + 0.004504205659031868, + 0.0053443084470927715, + 0.006088635418564081, + 0.005365161690860987, + 0.006358160637319088, + 0.005286499857902527, + 0.04576651751995087, + 0.04651998355984688, + 0.029893111437559128, + 0.049920398741960526, + 0.05170191451907158, + 0.04363739490509033, + 0.041500695049762726, + 0.038888148963451385, + 0.006975286640226841, + 0.007658281363546848, + 0.007058739196509123, + 0.008713741786777973, + 0.007899553515017033, + 0.007645873818546534, + 0.007005506195127964, + 0.005456168670207262, + 0.009245837107300758, + 0.007544889580458403, + 0.011745084077119827, + 0.006086323875933886, + 0.005556290969252586, + 0.009371347725391388, + 0.006293967831879854, + 0.007551627233624458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.020613020285964012 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193040043115616 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193040043115616, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193040043115616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10677077621221542, + 0.15439115464687347, + 0.06815342605113983, + 0.07851646840572357, + 0.19410060346126556, + 0.04484570398926735, + 0.04640335589647293, + 0.197614848613739, + 0.07166530191898346, + 0.03753840923309326 + ], + "credit_map": [ + 0.01556990947574377, + 0.01934187300503254, + 0.010171000845730305, + 0.013160651549696922, + 0.020813101902604103, + 0.011449046432971954, + 0.009949935600161552, + 0.006315259728580713, + 0.0088880006223917, + 0.023348230868577957, + 0.007715015672147274, + 0.006682353559881449, + 0.023801323026418686, + 0.059670235961675644, + 0.0064413826912641525, + 0.01784462109208107, + 0.005473677068948746, + 0.006333422847092152, + 0.006036390084773302, + 0.013348481617867947, + 0.009064755402505398, + 0.006107758730649948, + 0.01196409109979868, + 0.009824852459132671, + 0.004161218646913767, + 0.010078554041683674, + 0.004436600022017956, + 0.020028192549943924, + 0.0074127246625721455, + 0.0048850965686142445, + 0.01154207345098257, + 0.015972012653946877, + 0.010860278271138668, + 0.015823399648070335, + 0.008201484568417072, + 0.05546807497739792, + 0.017010493203997612, + 0.06784328818321228, + 0.008610106073319912, + 0.010283474810421467, + 0.005258963909000158, + 0.004695496987551451, + 0.0042634569108486176, + 0.006601657252758741, + 0.0038113207556307316, + 0.004770363215357065, + 0.008252800442278385, + 0.007191644515842199, + 0.007026216015219688, + 0.006682892329990864, + 0.004343728069216013, + 0.004753166809678078, + 0.005929989740252495, + 0.004645504988729954, + 0.007149545941501856, + 0.005872312467545271, + 0.0274006687104702, + 0.024576496332883835, + 0.01798395812511444, + 0.02799156680703163, + 0.03253377228975296, + 0.024667540565133095, + 0.023048199713230133, + 0.019412653520703316, + 0.009335525333881378, + 0.010220238007605076, + 0.006421807687729597, + 0.007389052305370569, + 0.013157176785171032, + 0.009262405335903168, + 0.007422290276736021, + 0.008456806652247906, + 0.006544190458953381, + 0.0057052732445299625, + 0.002992980182170868, + 0.0046298922970891, + 0.00403244374319911, + 0.0064426749013364315, + 0.00435181288048625, + 0.002839143853634596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.018193040043115616 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.051222119480371475, + 0.05755224823951721, + 0.05271356552839279, + 0.07173993438482285, + 0.2766944468021393, + 0.15874968469142914, + 0.05960381403565407, + 0.060398001223802567, + 0.051966335624456406, + 0.1593598574399948 + ], + "credit_map": [ + 0.0053538186475634575, + 0.0074982549995183945, + 0.005540529265999794, + 0.00884401984512806, + 0.0071537126787006855, + 0.006356659810990095, + 0.0064968834631145, + 0.003978235647082329, + 0.00461344001814723, + 0.009816121309995651, + 0.005470443982630968, + 0.002717384370043874, + 0.016202695667743683, + 0.010401510633528233, + 0.004642115905880928, + 0.0036885335575789213, + 0.003578231669962406, + 0.0074476501904428005, + 0.0038858363404870033, + 0.006682486739009619, + 0.008167157880961895, + 0.0046677859500050545, + 0.012454507872462273, + 0.005829907022416592, + 0.006370791234076023, + 0.010394936427474022, + 0.005173754412680864, + 0.01330653391778469, + 0.007503605913370848, + 0.008749563246965408, + 0.008675020188093185, + 0.011565729044377804, + 0.013458467088639736, + 0.029559403657913208, + 0.015166882425546646, + 0.06856679916381836, + 0.038125742226839066, + 0.08241351693868637, + 0.01720183528959751, + 0.012201780453324318, + 0.021715302020311356, + 0.02630305476486683, + 0.01460260059684515, + 0.020029302686452866, + 0.018336204811930656, + 0.01690870150923729, + 0.018149634823203087, + 0.02270488627254963, + 0.008083196356892586, + 0.006420259363949299, + 0.005201683845371008, + 0.00744536891579628, + 0.005951590836048126, + 0.006474348250776529, + 0.010649285279214382, + 0.009378071874380112, + 0.011389707215130329, + 0.006896467879414558, + 0.005659902468323708, + 0.007048974744975567, + 0.008029280230402946, + 0.008833184838294983, + 0.006623219233006239, + 0.005917256698012352, + 0.005841380450874567, + 0.006651198025792837, + 0.004806261043995619, + 0.007898637093603611, + 0.00563573045656085, + 0.006101491395384073, + 0.00876779854297638, + 0.006263834424316883, + 0.016997843980789185, + 0.014509179629385471, + 0.02839481644332409, + 0.019689463078975677, + 0.015392270870506763, + 0.021984299644827843, + 0.02152697555720806, + 0.02086498774588108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.023285161703824997 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05336616188287735, + 0.15551966428756714, + 0.04172508418560028, + 0.02944403886795044, + 0.03657662868499756, + 0.08226995170116425, + 0.03826246038079262, + 0.07146265357732773, + 0.05516664683818817, + 0.4362066984176636 + ], + "credit_map": [ + 0.006739396136254072, + 0.009048199281096458, + 0.00542542664334178, + 0.007037182804197073, + 0.007487492170184851, + 0.007355579175055027, + 0.006048556882888079, + 0.004224328324198723, + 0.007592995651066303, + 0.026287227869033813, + 0.006907845381647348, + 0.007706096861511469, + 0.028359854593873024, + 0.05818759277462959, + 0.011665930040180683, + 0.008812131360173225, + 0.0033968312200158834, + 0.004571634344756603, + 0.004345903638750315, + 0.008597789332270622, + 0.004195735324174166, + 0.0036259936168789864, + 0.007688684388995171, + 0.005302508361637592, + 0.0021119241137057543, + 0.0046612778678536415, + 0.0017245649360120296, + 0.007074149325489998, + 0.002653270959854126, + 0.003128794953227043, + 0.0030263494700193405, + 0.005063704680651426, + 0.0033688389230519533, + 0.006408111657947302, + 0.002639731392264366, + 0.006453972775489092, + 0.005068361293524504, + 0.00734470272436738, + 0.00255516916513443, + 0.0027377395890653133, + 0.011047065258026123, + 0.00753185898065567, + 0.007880745455622673, + 0.011764389462769032, + 0.006050366908311844, + 0.01163552887737751, + 0.013872657902538776, + 0.01248733326792717, + 0.005536428652703762, + 0.005574784241616726, + 0.0029129069298505783, + 0.0041711884550750256, + 0.0043715075589716434, + 0.004544810391962528, + 0.006482859142124653, + 0.004667976871132851, + 0.010868806391954422, + 0.00826975330710411, + 0.006242417264729738, + 0.01087028905749321, + 0.006968000903725624, + 0.009748639538884163, + 0.011241898871958256, + 0.007252844516187906, + 0.007367119658738375, + 0.00597963435575366, + 0.005227175075560808, + 0.004948294721543789, + 0.0076592788100242615, + 0.007502274587750435, + 0.005691096652299166, + 0.010791770182549953, + 0.04949966445565224, + 0.03931468352675438, + 0.08276370912790298, + 0.04726535826921463, + 0.0408407598733902, + 0.0672333613038063, + 0.05199326202273369, + 0.05729590728878975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.020579971373081207 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, overhead close-up, the camera slowly zooms in, closing in on the behemoth Godzilla. Under a pitch-black night sky, shrouded in lightning and thunder, it stands amid the rubble of a city, surrounded by twisted steel and shattered concrete. The air is thick with smoke and ash. Godzilla's armor, clad in deep black scales as rough and battle-worn as ancient rock, stares forward with eyes blazing with atomic blue fury. A maw studded with dagger-sharp teeth suddenly opens, and a deafening roar erupts from its deep throat. Its massive claws loom menacingly at the edge of the frame. The force of the sonic boom distorts the air, sending rubble from the crumbling buildings around it careening down. At that moment, a sharp blue anamorphic lens flare slowly and gracefully streaks across its scarred, hideous face, briefly illuminating the rough texture of its skin and the utter fury of its expression. The entire scene is illuminated by distant city flames and occasional flashes of lightning, creating an apocalyptic, tense atmosphere filled with destructive power and primal rage.", + "seed": 10996, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10996_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": -0.18225132850930095, + "vlm_logprob": -0.18225132850930095, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.18225132850930095, + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6561520099639893 + ], + "opposite_logprob_tokens": [ + -0.7811520099639893 + ], + "logprob_reward": -0.6561520099639893, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6561520099639893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32593899965286255, + 0.2523343861103058, + 0.024303218349814415, + 0.04272541403770447, + 0.0573650486767292, + 0.04701418802142143, + 0.06195717304944992, + 0.05394597724080086, + 0.10260460525751114, + 0.03181099519133568 + ], + "credit_map": [ + 0.035384319722652435, + 0.030161680653691292, + 0.03016483783721924, + 0.05714866518974304, + 0.045229651033878326, + 0.0421968512237072, + 0.06110409274697304, + 0.02454889751970768, + 0.03941149264574051, + 0.02983740344643593, + 0.02965797483921051, + 0.02873952127993107, + 0.028886722400784492, + 0.037909433245658875, + 0.03170463815331459, + 0.026187190786004066, + 0.002534580649808049, + 0.003022263990715146, + 0.0035475564654916525, + 0.0037640437949448824, + 0.003891232656314969, + 0.0032020516227930784, + 0.002369751688092947, + 0.0019717386458069086, + 0.004977221600711346, + 0.005237261764705181, + 0.005415935069322586, + 0.007034363690763712, + 0.005010612308979034, + 0.004823010880500078, + 0.004956316202878952, + 0.005270694382488728, + 0.010893560945987701, + 0.005553929600864649, + 0.006696303375065327, + 0.004809842444956303, + 0.007841856218874454, + 0.005509238224476576, + 0.006562090944498777, + 0.00949822273105383, + 0.005702346097677946, + 0.006045081652700901, + 0.0060042510740458965, + 0.0077928295359015465, + 0.00642906129360199, + 0.006279570981860161, + 0.004160167649388313, + 0.004600879270583391, + 0.005545012652873993, + 0.007144899573177099, + 0.00473018828779459, + 0.010168967768549919, + 0.007564387749880552, + 0.009266949258744717, + 0.006055758334696293, + 0.011481013149023056, + 0.0055572399869561195, + 0.009938708506524563, + 0.006826708093285561, + 0.008931173011660576, + 0.005620307754725218, + 0.005119523964822292, + 0.005563382059335709, + 0.006388938520103693, + 0.004501216113567352, + 0.004949287045747042, + 0.03945060446858406, + 0.03447685390710831, + 0.008259943686425686, + 0.0038270202931016684, + 0.004053138196468353, + 0.0030865350272506475, + 0.0016476541059091687, + 0.004133392591029406, + 0.004875692073255777, + 0.0031283057760447264, + 0.004710231442004442, + 0.005689059849828482, + 0.0029307380318641663, + 0.004695921204984188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.6561520099639893 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159946382045746 + ], + "opposite_logprob_tokens": [ + -1.7015994787216187 + ], + "logprob_reward": -0.20159946382045746, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159946382045746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15677201747894287, + 0.30942779779434204, + 0.03701338544487953, + 0.2550453245639801, + 0.026910047978162766, + 0.053530674427747726, + 0.056661494076251984, + 0.028956567868590355, + 0.04721499979496002, + 0.02846773900091648 + ], + "credit_map": [ + 0.016466107219457626, + 0.0117690684273839, + 0.030055822804570198, + 0.021292846649885178, + 0.01949002780020237, + 0.020022565498948097, + 0.01951751671731472, + 0.018158061429858208, + 0.047063153237104416, + 0.03163476660847664, + 0.03425607085227966, + 0.03398970142006874, + 0.03340661898255348, + 0.05355837196111679, + 0.04632773622870445, + 0.02919134870171547, + 0.00569074647501111, + 0.0043374584056437016, + 0.0054201846942305565, + 0.004433806519955397, + 0.004934564232826233, + 0.005104121286422014, + 0.0034603567328304052, + 0.0036321463994681835, + 0.022659601643681526, + 0.03292115405201912, + 0.022856561467051506, + 0.035235099494457245, + 0.03168513625860214, + 0.030039727687835693, + 0.02773193269968033, + 0.05191611498594284, + 0.00337113905698061, + 0.003422483103349805, + 0.003512253984808922, + 0.004051383584737778, + 0.003319505834951997, + 0.003050948493182659, + 0.0024004378356039524, + 0.0037818944547325373, + 0.005107298493385315, + 0.00558133190497756, + 0.005406074225902557, + 0.011731362901628017, + 0.0069593521766364574, + 0.006526292767375708, + 0.006802247371524572, + 0.0054167150519788265, + 0.004477235488593578, + 0.004592213314026594, + 0.002536651911213994, + 0.00601324625313282, + 0.014256063848733902, + 0.007346149068325758, + 0.010340024717152119, + 0.007099907845258713, + 0.0028922278434038162, + 0.005387965124100447, + 0.003776530735194683, + 0.004266631323844194, + 0.0025864255148917437, + 0.0024406216107308865, + 0.0028743597213178873, + 0.004731805995106697, + 0.00263042701408267, + 0.0026099944952875376, + 0.01425000000745058, + 0.005687842145562172, + 0.013878950849175453, + 0.003950673155486584, + 0.00231173075735569, + 0.001895384630188346, + 0.002055383287370205, + 0.004115609917789698, + 0.004387348890304565, + 0.004410404246300459, + 0.003133518621325493, + 0.002968603279441595, + 0.003619526280090213, + 0.0037773449439555407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.20159946382045746 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636283989995718 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636283989995718, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636283989995718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12462794035673141, + 0.2329091578722, + 0.03556228056550026, + 0.12848670780658722, + 0.10835017263889313, + 0.029175279662013054, + 0.09616415202617645, + 0.07323472946882248, + 0.13509827852249146, + 0.0363912358880043 + ], + "credit_map": [ + 0.01328166201710701, + 0.013955343514680862, + 0.01914173923432827, + 0.017496367916464806, + 0.015332036651670933, + 0.016117645427584648, + 0.015290436334908009, + 0.014012722298502922, + 0.03384585678577423, + 0.03450682386755943, + 0.028048327192664146, + 0.025916989892721176, + 0.030878232792019844, + 0.03139384090900421, + 0.024801788851618767, + 0.0235173087567091, + 0.0035394097212702036, + 0.004419275559484959, + 0.005580989643931389, + 0.0043290890753269196, + 0.0047189705073833466, + 0.005461298394948244, + 0.004159566946327686, + 0.0033536842092871666, + 0.013786627911031246, + 0.015993699431419373, + 0.012534708715975285, + 0.013988533057272434, + 0.014909131452441216, + 0.014140660874545574, + 0.015714041888713837, + 0.027419313788414, + 0.013903985731303692, + 0.01288970559835434, + 0.016205035150051117, + 0.01290845312178135, + 0.01596827432513237, + 0.009692489169538021, + 0.013081640936434269, + 0.013700591400265694, + 0.004278351552784443, + 0.004394141025841236, + 0.0030627730302512646, + 0.00462062144652009, + 0.0030745125841349363, + 0.0036509851925075054, + 0.0034983723890036345, + 0.002595525700598955, + 0.008560257963836193, + 0.008266839198768139, + 0.008210713975131512, + 0.017135776579380035, + 0.012916315346956253, + 0.014004289172589779, + 0.012960498221218586, + 0.014109469018876553, + 0.005778538063168526, + 0.011248434893786907, + 0.006523872260004282, + 0.011711379513144493, + 0.006022647954523563, + 0.012715833261609077, + 0.005000891629606485, + 0.014233134686946869, + 0.00651828246191144, + 0.004796497989445925, + 0.04778981953859329, + 0.024733521044254303, + 0.034446798264980316, + 0.0070457435213029385, + 0.005553440656512976, + 0.004214183427393436, + 0.001851145876571536, + 0.004194203764200211, + 0.006325745489448309, + 0.0033536111004650593, + 0.004863375797867775, + 0.007133694365620613, + 0.003276508767157793, + 0.005392956081777811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.007636283989995718 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606813259422779 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606813259422779, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606813259422779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18818916380405426, + 0.19785432517528534, + 0.028943445533514023, + 0.03292187303304672, + 0.0886995941400528, + 0.02787022665143013, + 0.2726414203643799, + 0.0509612150490284, + 0.08164174109697342, + 0.030276969075202942 + ], + "credit_map": [ + 0.018225545063614845, + 0.018999408930540085, + 0.02115369588136673, + 0.031644441187381744, + 0.02760804444551468, + 0.02233879268169403, + 0.030734559521079063, + 0.01748468540608883, + 0.031475577503442764, + 0.024908389896154404, + 0.02394978143274784, + 0.023837419226765633, + 0.021838467568159103, + 0.027612291276454926, + 0.023053446784615517, + 0.021178964525461197, + 0.003333216067403555, + 0.002840524772182107, + 0.00445446465164423, + 0.003776290686801076, + 0.004365378525108099, + 0.0028322022408246994, + 0.0029148967005312443, + 0.004426474217325449, + 0.003639657748863101, + 0.0035154831130057573, + 0.002468396909534931, + 0.003973032347857952, + 0.004162730183452368, + 0.004024139605462551, + 0.004145040642470121, + 0.006993396207690239, + 0.011000699363648891, + 0.010695220902562141, + 0.010836374014616013, + 0.012651344761252403, + 0.010945774614810944, + 0.008635248988866806, + 0.014023367315530777, + 0.00991157814860344, + 0.0032066369894891977, + 0.0040215118788182735, + 0.0035066904965788126, + 0.0050639621913433075, + 0.0031463552732020617, + 0.0037633131723850965, + 0.002360844984650612, + 0.002800911897793412, + 0.019001370295882225, + 0.026041194796562195, + 0.018466919660568237, + 0.03781721368432045, + 0.050982654094696045, + 0.03567657247185707, + 0.05218038335442543, + 0.03247511759400368, + 0.003170591779053211, + 0.005552978720515966, + 0.004988342057913542, + 0.007104835007339716, + 0.0037925741635262966, + 0.006505576893687248, + 0.00580989895388484, + 0.014036421664059162, + 0.005284592043608427, + 0.006477325689047575, + 0.017848722636699677, + 0.009948343969881535, + 0.02137828804552555, + 0.008289062418043613, + 0.006858667358756065, + 0.00555674871429801, + 0.001459835679270327, + 0.0024553993716835976, + 0.0046631102450191975, + 0.0029415928293019533, + 0.004630004055798054, + 0.0070220353081822395, + 0.0025459390599280596, + 0.004559055436402559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.01606813259422779 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980075217783451 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980075217783451, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980075217783451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13275623321533203, + 0.33627232909202576, + 0.02765045501291752, + 0.07827040553092957, + 0.023171475157141685, + 0.019149523228406906, + 0.07294711470603943, + 0.06709866970777512, + 0.16695547103881836, + 0.07572830468416214 + ], + "credit_map": [ + 0.012468360364437103, + 0.013813118450343609, + 0.01765696331858635, + 0.021024765446782112, + 0.015890777111053467, + 0.016260629519820213, + 0.018655572086572647, + 0.016986044123768806, + 0.05123210325837135, + 0.03536400943994522, + 0.0391080267727375, + 0.037441227585077286, + 0.035768862813711166, + 0.055273741483688354, + 0.048596471548080444, + 0.03348790481686592, + 0.004784448072314262, + 0.0030378224328160286, + 0.00410206476226449, + 0.004462327342480421, + 0.0029933538753539324, + 0.003521554172039032, + 0.0030203985515981913, + 0.001728482311591506, + 0.006711212452501059, + 0.010166828520596027, + 0.0074800592847168446, + 0.010093952529132366, + 0.009627965278923512, + 0.008683102205395699, + 0.009188150987029076, + 0.016319144517183304, + 0.003030245192348957, + 0.0025137951597571373, + 0.0030354191549122334, + 0.0030171223916113377, + 0.0030037136748433113, + 0.0030553219839930534, + 0.0025400780141353607, + 0.0029757781885564327, + 0.00234599388204515, + 0.0018695428734645247, + 0.0025431255344301462, + 0.0038198886904865503, + 0.0021103876642882824, + 0.002653545932844281, + 0.0018701126100495458, + 0.0019369255751371384, + 0.005193464457988739, + 0.010958396829664707, + 0.005526641383767128, + 0.00995020940899849, + 0.01168098859488964, + 0.0097023481503129, + 0.010295051150023937, + 0.009640014730393887, + 0.005892706569284201, + 0.006890880409628153, + 0.006823675241321325, + 0.008793937973678112, + 0.005748191382735968, + 0.009124755859375, + 0.007177087478339672, + 0.01664743572473526, + 0.007356893736869097, + 0.0081862211227417, + 0.05828569084405899, + 0.042666055262088776, + 0.024677712470293045, + 0.011107771657407284, + 0.007522610481828451, + 0.007152522914111614, + 0.0038231792859733105, + 0.008205322548747063, + 0.011562650091946125, + 0.006437493488192558, + 0.01202381867915392, + 0.017185578122735023, + 0.0056715914979577065, + 0.010818674229085445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.02980075217783451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a powerful, overhead close-up, the camera slowly zooms in, closing in on the behemoth Godzilla. Under a pitch-black night sky, shrouded in lightning and thunder, it stands amid the rubble of a city, surrounded by twisted steel and shattered concrete. The air is thick with smoke and ash. Godzilla's armor, clad in deep black scales as rough and battle-worn as ancient rock, stares forward with eyes blazing with atomic blue fury. A maw studded with dagger-sharp teeth suddenly opens, and a deafening roar erupts from its deep throat. Its massive claws loom menacingly at the edge of the frame. The force of the sonic boom distorts the air, sending rubble from the crumbling buildings around it careening down. At that moment, a sharp blue anamorphic lens flare slowly and gracefully streaks across its scarred, hideous face, briefly illuminating the rough texture of its skin and the utter fury of its expression. The entire scene is illuminated by distant city flames and occasional flashes of lightning, creating an apocalyptic, tense atmosphere filled with destructive power and primal rage.", + "seed": 10997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_7/index_3069_seed_10997_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": -0.08859107354655862, + "vlm_logprob": -0.08859107354655862, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.08859107354655862, + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32610124349594116 + ], + "opposite_logprob_tokens": [ + -1.326101303100586 + ], + "logprob_reward": -0.32610124349594116, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32610124349594116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032723937183618546, + 0.0784405842423439, + 0.26966628432273865, + 0.19147329032421112, + 0.10678129643201828, + 0.12188175320625305, + 0.034241318702697754, + 0.037082552909851074, + 0.035238172858953476, + 0.09247075021266937 + ], + "credit_map": [ + 0.003965888638049364, + 0.0041494290344417095, + 0.006067798938602209, + 0.002541924361139536, + 0.00498114712536335, + 0.005220259539783001, + 0.002717097755521536, + 0.003080393187701702, + 0.009889976121485233, + 0.01815313845872879, + 0.008520552888512611, + 0.007945330813527107, + 0.012762904167175293, + 0.008881130255758762, + 0.006250657606869936, + 0.006036902312189341, + 0.013039919547736645, + 0.015331712551414967, + 0.04262677580118179, + 0.02957242913544178, + 0.03928731754422188, + 0.06465929746627808, + 0.0388079434633255, + 0.026340896263718605, + 0.020176701247692108, + 0.021712491288781166, + 0.020666668191552162, + 0.03528744354844093, + 0.022023938596248627, + 0.029278459027409554, + 0.024258537217974663, + 0.018069034442305565, + 0.014986302703619003, + 0.007244196720421314, + 0.016652483493089676, + 0.011548726819455624, + 0.012414565309882164, + 0.020872510969638824, + 0.010221880860626698, + 0.012840638868510723, + 0.01857621781527996, + 0.013803963549435139, + 0.012883646413683891, + 0.01211938913911581, + 0.01226831041276455, + 0.023998090997338295, + 0.013936597853899002, + 0.014295527711510658, + 0.004421199206262827, + 0.004404495935887098, + 0.004694393835961819, + 0.005532620474696159, + 0.003739573061466217, + 0.0033512613736093044, + 0.0036812624894082546, + 0.004416515119373798, + 0.004767462611198425, + 0.004094294738024473, + 0.0047467537224292755, + 0.007232240866869688, + 0.0038779699243605137, + 0.003463667817413807, + 0.005545766558498144, + 0.0033543978352099657, + 0.004419079981744289, + 0.004591604694724083, + 0.0038153629284352064, + 0.004921928979456425, + 0.004624429624527693, + 0.004426469095051289, + 0.003969324752688408, + 0.0044699725694954395, + 0.007114442065358162, + 0.008064284920692444, + 0.014398554340004921, + 0.019065644592046738, + 0.008663171902298927, + 0.007809168193489313, + 0.01249262597411871, + 0.014862862415611744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.32610124349594116 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905833214521408 + ], + "opposite_logprob_tokens": [ + -2.4640583992004395 + ], + "logprob_reward": -0.08905833214521408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905833214521408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03644167631864548, + 0.2379520833492279, + 0.35820525884628296, + 0.024857591837644577, + 0.02080783061683178, + 0.09398970752954483, + 0.04406386613845825, + 0.02403385564684868, + 0.0882158875465393, + 0.07143223285675049 + ], + "credit_map": [ + 0.0064744120463728905, + 0.004586115013808012, + 0.008850658312439919, + 0.0027429345063865185, + 0.0036573081742972136, + 0.004706165753304958, + 0.0027193042915314436, + 0.002704779151827097, + 0.02877791039645672, + 0.060669247061014175, + 0.01901104487478733, + 0.024877650663256645, + 0.03881380707025528, + 0.03218386322259903, + 0.018538599833846092, + 0.015079949982464314, + 0.01631210558116436, + 0.01910160668194294, + 0.05578954517841339, + 0.02693529985845089, + 0.0535329133272171, + 0.10374805331230164, + 0.051893871277570724, + 0.03089185804128647, + 0.003004583530128002, + 0.002178254770115018, + 0.003291107714176178, + 0.005888649728149176, + 0.002911492483690381, + 0.0025202324613928795, + 0.0030334494076669216, + 0.002029821975156665, + 0.0038066208362579346, + 0.0017415785696357489, + 0.002212331397458911, + 0.0017748107202351093, + 0.003051867475733161, + 0.002808950375765562, + 0.0030489619821310043, + 0.002362707396969199, + 0.013038596138358116, + 0.010380568914115429, + 0.01524080615490675, + 0.01057275291532278, + 0.011158875189721584, + 0.013092032633721828, + 0.011671807616949081, + 0.008834262378513813, + 0.00704925786703825, + 0.0046160779893398285, + 0.00638975715264678, + 0.004816961474716663, + 0.004528435412794352, + 0.004248140379786491, + 0.00718723563477397, + 0.005227998364716768, + 0.003219885751605034, + 0.0020770286209881306, + 0.003965526353567839, + 0.004029368981719017, + 0.002256933832541108, + 0.0015359669923782349, + 0.00395053718239069, + 0.0029986086301505566, + 0.012487374246120453, + 0.00996249821037054, + 0.011563192121684551, + 0.013188469223678112, + 0.009993577376008034, + 0.009903205558657646, + 0.01179959625005722, + 0.009317975491285324, + 0.005834061186760664, + 0.00608806824311614, + 0.010994772426784039, + 0.013927971944212914, + 0.006983296945691109, + 0.005920902825891972, + 0.010549400933086872, + 0.011133763939142227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.08905833214521408 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632735185325146 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632735185325146, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632735185325146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0316946916282177, + 0.0842755064368248, + 0.39384958148002625, + 0.15525299310684204, + 0.13856758177280426, + 0.09458785504102707, + 0.017398374155163765, + 0.026264239102602005, + 0.02692110277712345, + 0.03118807077407837 + ], + "credit_map": [ + 0.0039198389276862144, + 0.005069933831691742, + 0.005872718058526516, + 0.0027909581549465656, + 0.004311611410230398, + 0.004335523117333651, + 0.002312316093593836, + 0.0030817922670394182, + 0.008309061639010906, + 0.0146532217040658, + 0.011451172642409801, + 0.007762235589325428, + 0.01281038112938404, + 0.010968733578920364, + 0.011156829074025154, + 0.007163866423070431, + 0.019030846655368805, + 0.02047787606716156, + 0.05793941020965576, + 0.031028013676404953, + 0.060999348759651184, + 0.10953113436698914, + 0.057321708649396896, + 0.03752125799655914, + 0.007753791753202677, + 0.016690680757164955, + 0.023391688242554665, + 0.04232925549149513, + 0.017285233363509178, + 0.018404338508844376, + 0.017011815682053566, + 0.012386192567646503, + 0.02009018510580063, + 0.011498695239424706, + 0.020819956436753273, + 0.014463730156421661, + 0.015035944990813732, + 0.02516496181488037, + 0.015208938159048557, + 0.016285179182887077, + 0.011586486361920834, + 0.010452906601130962, + 0.015587324276566505, + 0.006096867844462395, + 0.013191759586334229, + 0.013097289018332958, + 0.014618009328842163, + 0.009957210160791874, + 0.002427969593554735, + 0.001793367206119001, + 0.00204131449572742, + 0.0019667833112180233, + 0.002522130962461233, + 0.0023616987746208906, + 0.002221675356850028, + 0.0020634341053664684, + 0.0036659869365394115, + 0.00315910205245018, + 0.0032789278775453568, + 0.0034634280018508434, + 0.003048664191737771, + 0.0024431704077869654, + 0.004123363643884659, + 0.0030815955251455307, + 0.0032574355136603117, + 0.003103179158642888, + 0.003348367288708687, + 0.004740653093904257, + 0.00290648709051311, + 0.002871520584449172, + 0.003624530043452978, + 0.0030689274426549673, + 0.0026349627878516912, + 0.00359724392183125, + 0.004623145796358585, + 0.005093231797218323, + 0.0029836916364729404, + 0.002774776192381978, + 0.004605314694344997, + 0.00487570371478796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.007632735185325146 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03527728095650673, + 0.14542876183986664, + 0.1580205261707306, + 0.09086550772190094, + 0.0957283228635788, + 0.07794944196939468, + 0.03389445319771767, + 0.14435015618801117, + 0.13508614897727966, + 0.08339940011501312 + ], + "credit_map": [ + 0.00544514786452055, + 0.004117204807698727, + 0.007839670404791832, + 0.0033221114426851273, + 0.00369516690261662, + 0.005658397916704416, + 0.0028376742266118526, + 0.0023619052954018116, + 0.016995739191770554, + 0.03606916218996048, + 0.011405770666897297, + 0.013274899683892727, + 0.02318591997027397, + 0.021967865526676178, + 0.014079960994422436, + 0.008449427783489227, + 0.007903160527348518, + 0.019281862303614616, + 0.010975953191518784, + 0.016409866511821747, + 0.02815556712448597, + 0.042897872626781464, + 0.019759293645620346, + 0.01263693068176508, + 0.005513872019946575, + 0.008881870657205582, + 0.013556589372456074, + 0.024664098396897316, + 0.00996167492121458, + 0.010911152698099613, + 0.009711903519928455, + 0.007664342410862446, + 0.013625388033688068, + 0.006829859223216772, + 0.015826279297471046, + 0.009739229455590248, + 0.010933274403214455, + 0.01773105189204216, + 0.00950663536787033, + 0.011536591686308384, + 0.009991777129471302, + 0.009995799511671066, + 0.00843773502856493, + 0.010665275156497955, + 0.010431298986077309, + 0.010252305306494236, + 0.010383720509707928, + 0.007791530806571245, + 0.0050822664052248, + 0.002845596754923463, + 0.004250532481819391, + 0.0045161196030676365, + 0.0036828231532126665, + 0.0034190400037914515, + 0.005223846063017845, + 0.004874227102845907, + 0.01872210204601288, + 0.014455139636993408, + 0.017492394894361496, + 0.032484859228134155, + 0.012291783466935158, + 0.00831510964781046, + 0.024281028658151627, + 0.016307728365063667, + 0.016821671277284622, + 0.017747504636645317, + 0.01743433251976967, + 0.03306421637535095, + 0.010394598357379436, + 0.009534955024719238, + 0.014536862261593342, + 0.015551996417343616, + 0.006628705654293299, + 0.007931243628263474, + 0.012581474147737026, + 0.01596078835427761, + 0.007885894738137722, + 0.006610440090298653, + 0.012282985262572765, + 0.013517859391868114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.004096213262528181 + }, + { + "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 in an overhead close-up view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom in on Godzilla?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Godzilla's skin depicted as deep black scales with a rough, battle-worn texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Godzilla open its maw to reveal dagger-sharp teeth and emit a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue anamorphic lens flare visible streaking across Godzilla's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0771065354347229, + 0.10719439387321472, + 0.17657342553138733, + 0.12720029056072235, + 0.1289796531200409, + 0.08239227533340454, + 0.015813209116458893, + 0.0390118770301342, + 0.06281758844852448, + 0.1829107701778412 + ], + "credit_map": [ + 0.010079938918352127, + 0.01037288736552, + 0.011974985711276531, + 0.008009780198335648, + 0.012256743386387825, + 0.012368847616016865, + 0.007368880789726973, + 0.004674466326832771, + 0.01051468588411808, + 0.02264513447880745, + 0.008221974596381187, + 0.013557996600866318, + 0.01552977692335844, + 0.01650087535381317, + 0.01203735638409853, + 0.008186588063836098, + 0.009213675744831562, + 0.019263917580246925, + 0.014395175501704216, + 0.0166375320404768, + 0.030755693092942238, + 0.04943951219320297, + 0.02289017103612423, + 0.013977750204503536, + 0.011770852841436863, + 0.013414795510470867, + 0.009178347885608673, + 0.01679656282067299, + 0.02288491278886795, + 0.019708113744854927, + 0.020186100155115128, + 0.0132606141269207, + 0.01876600831747055, + 0.00940531212836504, + 0.020816458389163017, + 0.014941904693841934, + 0.014078769832849503, + 0.02307191863656044, + 0.012767374515533447, + 0.015131901949644089, + 0.015050601214170456, + 0.005722885485738516, + 0.011544509790837765, + 0.006170195993036032, + 0.006176849827170372, + 0.01948155090212822, + 0.009329489432275295, + 0.008916190825402737, + 0.0025335769169032574, + 0.001211363822221756, + 0.0032113988418132067, + 0.0017898882506415248, + 0.0015666537219658494, + 0.0016001691110432148, + 0.0019366658525541425, + 0.0019634936470538378, + 0.005326598882675171, + 0.0032193006481975317, + 0.006803642492741346, + 0.005733237601816654, + 0.0036266238894313574, + 0.004005672410130501, + 0.005002057645469904, + 0.0052947429940104485, + 0.007104919292032719, + 0.008404326625168324, + 0.008738188073039055, + 0.010955999605357647, + 0.007334843743592501, + 0.006190203130245209, + 0.007223315071314573, + 0.006865796633064747, + 0.017699996009469032, + 0.017098408192396164, + 0.025853337720036507, + 0.030828891322016716, + 0.018529968336224556, + 0.015102176927030087, + 0.026650445535779, + 0.03114754892885685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.016066843643784523 + } + ], + "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/0000167/group_8/index_tensor([1070, 1070])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_tensor([1070, 1070])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ea276ca0c2abcb3af657cc2de509065310403c6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_tensor([1070, 1070])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, swooping shot begins at a gentle, high angle, then slowly descends and advances, gently following the movements of a delicate pansy. The flower is in full bloom, its velvety purple and yellow petals, embellished with delicate dark veins, standing out against a carpet of thick, emerald green, and moist moss. A soft, warm yellow light streams down from above and sideways, penetrating the edges of the petals, giving them a translucent quality and casting a dappled, warm glow across the moss. Under the lens's gaze, the pansy's stem delicately descends into the soft moss with an almost imperceptible, graceful rhythm. The green moss fibers gently part, welcoming the flower, like the meticulous construction of a cozy nest. A shallow depth of field gently blurs the background, focusing all attention on this microcosmic ritual of life, creating an atmosphere of profound tranquility, healing, and vibrant warmth.", + "seed": 2082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_1070_seed_2082_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": -0.006796169630251825, + "vlm_logprob": -0.006796169630251825, + "vlm_yes_no_margin": 5.549999952316284, + "vlm_token_credit": -0.006796169630251825, + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002148464322090149 + ], + "opposite_logprob_tokens": [ + -6.502148628234863 + ], + "logprob_reward": -0.002148464322090149, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.002148464322090149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14284801483154297, + 0.06807240843772888, + 0.039550818502902985, + 0.34184709191322327, + 0.06746707111597061, + 0.08416426926851273, + 0.036386992782354355, + 0.05229964479804039, + 0.08144328743219376, + 0.08592033386230469 + ], + "credit_map": [ + 0.009306971915066242, + 0.01772097870707512, + 0.01741786114871502, + 0.026575488969683647, + 0.014645655639469624, + 0.015407782047986984, + 0.021180972456932068, + 0.020592326298356056, + 0.0065306564792990685, + 0.004544827155768871, + 0.010475478135049343, + 0.008804556913673878, + 0.009821536019444466, + 0.013004952110350132, + 0.006255132611840963, + 0.008635278791189194, + 0.003909900318831205, + 0.004103177692741156, + 0.007022303994745016, + 0.007204495836049318, + 0.0043947091326117516, + 0.00469877477735281, + 0.004861762281507254, + 0.0033556984271854162, + 0.008911829441785812, + 0.016440022736787796, + 0.008042367175221443, + 0.009813894517719746, + 0.1322898119688034, + 0.1399516612291336, + 0.015207866206765175, + 0.011189677752554417, + 0.006650661118328571, + 0.008133447729051113, + 0.0110022546723485, + 0.00819320511072874, + 0.005128281656652689, + 0.0086889136582613, + 0.006452156696468592, + 0.013218157924711704, + 0.00986186508089304, + 0.008257700130343437, + 0.014256907626986504, + 0.013867896981537342, + 0.009470597840845585, + 0.010442017577588558, + 0.0063130855560302734, + 0.011694211512804031, + 0.0029886469710618258, + 0.0032702446915209293, + 0.0052399542182683945, + 0.006381344981491566, + 0.003777996404096484, + 0.004676843527704477, + 0.00556077528744936, + 0.004491191823035479, + 0.005910402629524469, + 0.006419472396373749, + 0.00871985126286745, + 0.006925743073225021, + 0.005956893786787987, + 0.005636235233396292, + 0.006067303940653801, + 0.0066637503914535046, + 0.010084317997097969, + 0.013019760139286518, + 0.013859529979526997, + 0.007821209728717804, + 0.009875456802546978, + 0.007429993245750666, + 0.009361980482935905, + 0.009991052560508251, + 0.009548150934278965, + 0.010443913750350475, + 0.010169981978833675, + 0.014316658489406109, + 0.013305393978953362, + 0.0096342284232378, + 0.011376515962183475, + 0.0071254996582865715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.002148464322090149 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04467032104730606, + 0.032268378883600235, + 0.027790362015366554, + 0.3615013659000397, + 0.022098351269960403, + 0.04232415929436684, + 0.036944109946489334, + 0.05379766598343849, + 0.20212455093860626, + 0.17648077011108398 + ], + "credit_map": [ + 0.004358895123004913, + 0.006463758181780577, + 0.005057739559561014, + 0.005005448125302792, + 0.004796015564352274, + 0.004798786714673042, + 0.0063769035041332245, + 0.00781277846544981, + 0.0028572650626301765, + 0.002358588855713606, + 0.005132531281560659, + 0.004208753816783428, + 0.003941678442060947, + 0.005206201691180468, + 0.0037051280960440636, + 0.004858228377997875, + 0.0026745074428617954, + 0.0010332628153264523, + 0.0017399246571585536, + 0.001677731634117663, + 0.008625208400189877, + 0.007985066622495651, + 0.0019627148285508156, + 0.0020919463131576777, + 0.0068017831072211266, + 0.015111532062292099, + 0.005471228621900082, + 0.005010121501982212, + 0.1419479101896286, + 0.16362713277339935, + 0.008966774679720402, + 0.014564888551831245, + 0.0018889307975769043, + 0.0025819074362516403, + 0.003232568269595504, + 0.0030503333546221256, + 0.002368378918617964, + 0.003052760846912861, + 0.0026933408807963133, + 0.0032301312312483788, + 0.003683309070765972, + 0.003660369198769331, + 0.006496557965874672, + 0.006088131107389927, + 0.005310053937137127, + 0.005882809404283762, + 0.004250128287822008, + 0.0069528017193078995, + 0.003015767317265272, + 0.0021580331958830357, + 0.006077653728425503, + 0.004778942093253136, + 0.0023533981293439865, + 0.006637304089963436, + 0.006622121669352055, + 0.005300887860357761, + 0.00482791056856513, + 0.004390474408864975, + 0.007859663106501102, + 0.009248943068087101, + 0.003167428309097886, + 0.007184877991676331, + 0.009061689488589764, + 0.008056674152612686, + 0.01870913803577423, + 0.029413366690278053, + 0.03708808869123459, + 0.021001366898417473, + 0.017234642058610916, + 0.025267168879508972, + 0.02439279668033123, + 0.029018007218837738, + 0.01839803345501423, + 0.01532349456101656, + 0.026526322588324547, + 0.018093222752213478, + 0.013743460178375244, + 0.03543670475482941, + 0.029948409646749496, + 0.0190111231058836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.00863910373300314 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.27321338653564453, + 0.06443437188863754, + 0.05060601606965065, + 0.25589537620544434, + 0.06166684255003929, + 0.05881359428167343, + 0.041778091341257095, + 0.02503715083003044, + 0.06022341176867485, + 0.10833177715539932 + ], + "credit_map": [ + 0.019923392683267593, + 0.04144815728068352, + 0.020923877134919167, + 0.02385820634663105, + 0.052719596773386, + 0.03803236410021782, + 0.029618654400110245, + 0.04668913409113884, + 0.0052500260062515736, + 0.0054979571141302586, + 0.010681992396712303, + 0.009336084127426147, + 0.007244093343615532, + 0.007856789976358414, + 0.00897339079529047, + 0.009594044648110867, + 0.002207888290286064, + 0.0021015747915953398, + 0.0034315302036702633, + 0.003851300338283181, + 0.016083842143416405, + 0.015709972009062767, + 0.00385265308432281, + 0.0033672566059976816, + 0.005086302291601896, + 0.011002393439412117, + 0.004999218042939901, + 0.005701694171875715, + 0.09892036020755768, + 0.11297708749771118, + 0.006943652871996164, + 0.010264644399285316, + 0.00524314446374774, + 0.006368366070091724, + 0.010005502961575985, + 0.006983211264014244, + 0.0059449514374136925, + 0.012116012163460255, + 0.0045865681022405624, + 0.01041908748447895, + 0.0041190218180418015, + 0.004708298947662115, + 0.009160994552075863, + 0.010932120494544506, + 0.004609393887221813, + 0.008496828377246857, + 0.007340307347476482, + 0.009446630254387856, + 0.0028688625898212194, + 0.002941604470834136, + 0.00859217718243599, + 0.004724225029349327, + 0.002648592460900545, + 0.00797194428741932, + 0.005693997256457806, + 0.006336686667054892, + 0.0030263562221080065, + 0.0027485289610922337, + 0.003978528082370758, + 0.0031414648983627558, + 0.0026132240891456604, + 0.0030668347608298063, + 0.0030081237200647593, + 0.0034540884662419558, + 0.0064397091045975685, + 0.009299900382757187, + 0.008825568482279778, + 0.005571694578975439, + 0.00754618039354682, + 0.007990692742168903, + 0.006662243977189064, + 0.007887424901127815, + 0.010283012874424458, + 0.008950170129537582, + 0.018469953909516335, + 0.01590220257639885, + 0.01279723271727562, + 0.014089589938521385, + 0.017384089529514313, + 0.010455526411533356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.004094195086508989 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16225452721118927, + 0.08369626104831696, + 0.03325214609503746, + 0.04316921904683113, + 0.12776151299476624, + 0.0956689864397049, + 0.05713661015033722, + 0.052694275975227356, + 0.16915032267570496, + 0.17521610856056213 + ], + "credit_map": [ + 0.010624529793858528, + 0.02233261615037918, + 0.022195592522621155, + 0.02651379443705082, + 0.019893277436494827, + 0.012044019997119904, + 0.02020370587706566, + 0.02844698540866375, + 0.007006924133747816, + 0.00809151865541935, + 0.013328379951417446, + 0.012345199473202229, + 0.00922209769487381, + 0.00891229696571827, + 0.012166454456746578, + 0.012623393908143044, + 0.0028813693206757307, + 0.0036382824182510376, + 0.0030665192753076553, + 0.003213234944269061, + 0.006740886718034744, + 0.006594538222998381, + 0.00437290221452713, + 0.0027444143779575825, + 0.004340155981481075, + 0.004551428370177746, + 0.0051934076473116875, + 0.006201783660799265, + 0.004248361103236675, + 0.006025225855410099, + 0.006038210354745388, + 0.0065706465393304825, + 0.013286895118653774, + 0.012608464807271957, + 0.02714696153998375, + 0.01564476266503334, + 0.00785910151898861, + 0.020045598968863487, + 0.010048364289104939, + 0.02112136036157608, + 0.010970531962811947, + 0.011834701523184776, + 0.01504160463809967, + 0.013907074928283691, + 0.011273670010268688, + 0.00931396521627903, + 0.012211221270263195, + 0.011116215027868748, + 0.0037048733793199062, + 0.004762409720569849, + 0.010527222417294979, + 0.0070356568321585655, + 0.005466251168400049, + 0.00809076614677906, + 0.009752947837114334, + 0.007796481251716614, + 0.004610139410942793, + 0.004456180613487959, + 0.009073534049093723, + 0.007857459597289562, + 0.004063303582370281, + 0.00729846628382802, + 0.00837423000484705, + 0.00696096196770668, + 0.02009199932217598, + 0.0236705020070076, + 0.02105189859867096, + 0.01243654265999794, + 0.02029654011130333, + 0.02224823459982872, + 0.02186749503016472, + 0.027487121522426605, + 0.013470564968883991, + 0.016026170924305916, + 0.025281626731157303, + 0.02503977157175541, + 0.02188878506422043, + 0.02504589594900608, + 0.033526964485645294, + 0.014936322346329689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0009190387791022658 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818004623055458 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01818004623055458, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818004623055458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04459928348660469, + 0.056880611926317215, + 0.04373826086521149, + 0.21189182996749878, + 0.06817720085382462, + 0.036107126623392105, + 0.039632752537727356, + 0.05415990203619003, + 0.29283711314201355, + 0.15197589993476868 + ], + "credit_map": [ + 0.003592626191675663, + 0.005460604093968868, + 0.006539247464388609, + 0.00681223813444376, + 0.004259007517248392, + 0.004551092628389597, + 0.005804145243018866, + 0.007580322213470936, + 0.004702226724475622, + 0.004978897515684366, + 0.01043946947902441, + 0.0074497731402516365, + 0.0074650803580880165, + 0.00865157414227724, + 0.005292806308716536, + 0.007900786586105824, + 0.0028781243599951267, + 0.0028514242731034756, + 0.003379581496119499, + 0.004174680449068546, + 0.01294575072824955, + 0.012230291031301022, + 0.003168204566463828, + 0.002110206987708807, + 0.004593022633343935, + 0.009942237287759781, + 0.004541713744401932, + 0.00435910327360034, + 0.08211448788642883, + 0.09180431067943573, + 0.006410103756934404, + 0.008126855827867985, + 0.009247343055903912, + 0.0069660223089158535, + 0.015492750331759453, + 0.007055272348225117, + 0.004753327928483486, + 0.008136081509292126, + 0.005824679974466562, + 0.01070172805339098, + 0.005407202523201704, + 0.004408697132021189, + 0.005142076872289181, + 0.0054932148195803165, + 0.004013031255453825, + 0.003403141163289547, + 0.004274694714695215, + 0.003965068142861128, + 0.0025633194018155336, + 0.0037437928840517998, + 0.006398202385753393, + 0.005088463891297579, + 0.0033689099363982677, + 0.007256795186549425, + 0.005414469633251429, + 0.00579879991710186, + 0.00545254722237587, + 0.0051548718474805355, + 0.0076034702360630035, + 0.008014674298465252, + 0.004729640204459429, + 0.006587566342204809, + 0.009178869426250458, + 0.007438262924551964, + 0.028894759714603424, + 0.05303575098514557, + 0.03310183808207512, + 0.023687805980443954, + 0.04576161876320839, + 0.03588187322020531, + 0.03584584593772888, + 0.0366276316344738, + 0.01894298568367958, + 0.012313994579017162, + 0.02931518293917179, + 0.01937905326485634, + 0.014215718023478985, + 0.018415015190839767, + 0.022429456934332848, + 0.016964484006166458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.01818004623055458 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, swooping shot begins at a gentle, high angle, then slowly descends and advances, gently following the movements of a delicate pansy. The flower is in full bloom, its velvety purple and yellow petals, embellished with delicate dark veins, standing out against a carpet of thick, emerald green, and moist moss. A soft, warm yellow light streams down from above and sideways, penetrating the edges of the petals, giving them a translucent quality and casting a dappled, warm glow across the moss. Under the lens's gaze, the pansy's stem delicately descends into the soft moss with an almost imperceptible, graceful rhythm. The green moss fibers gently part, welcoming the flower, like the meticulous construction of a cozy nest. A shallow depth of field gently blurs the background, focusing all attention on this microcosmic ritual of life, creating an atmosphere of profound tranquility, healing, and vibrant warmth.", + "seed": 2083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_1070_seed_2083_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": -0.007252647913992405, + "vlm_logprob": -0.007252647913992405, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.007252647913992405, + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038330908864736557 + ], + "opposite_logprob_tokens": [ + -6.003833293914795 + ], + "logprob_reward": -0.0038330908864736557, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0038330908864736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20343244075775146, + 0.09992143511772156, + 0.11723792552947998, + 0.046125296503305435, + 0.05946432054042816, + 0.0660647377371788, + 0.10733999311923981, + 0.11878529191017151, + 0.05686880275607109, + 0.12475985288619995 + ], + "credit_map": [ + 0.016998616978526115, + 0.02178299054503441, + 0.04106609523296356, + 0.018568167462944984, + 0.023609884083271027, + 0.03452147915959358, + 0.02477075345814228, + 0.02211443893611431, + 0.011649641208350658, + 0.015133817680180073, + 0.012154112569987774, + 0.012558680959045887, + 0.0088136475533247, + 0.015820356085896492, + 0.014405383728444576, + 0.009385796263813972, + 0.019397597759962082, + 0.014970885589718819, + 0.014867707155644894, + 0.00985537376254797, + 0.017057321965694427, + 0.02273775264620781, + 0.0102440370246768, + 0.00810724776238203, + 0.003297989722341299, + 0.004753954242914915, + 0.009306544438004494, + 0.004543872084468603, + 0.0036898194812238216, + 0.012204855680465698, + 0.003959177993237972, + 0.004369084723293781, + 0.006756005343049765, + 0.011460987851023674, + 0.007740826811641455, + 0.0071512083522975445, + 0.007405840326100588, + 0.004685845226049423, + 0.005337412003427744, + 0.008926192298531532, + 0.007264911662787199, + 0.012906922027468681, + 0.009791260585188866, + 0.006360289640724659, + 0.009147700853645802, + 0.009924502111971378, + 0.005860071629285812, + 0.004809081554412842, + 0.006431982386857271, + 0.005900589749217033, + 0.009019376710057259, + 0.005139796994626522, + 0.04031577706336975, + 0.030200893059372902, + 0.005360583309084177, + 0.004970990587025881, + 0.01650993898510933, + 0.013404746539890766, + 0.017256522551178932, + 0.014335259795188904, + 0.013450044207274914, + 0.014802038669586182, + 0.016668103635311127, + 0.012358644045889378, + 0.007390179205685854, + 0.00448182038962841, + 0.009303008206188679, + 0.006141190882772207, + 0.008554146625101566, + 0.005760649219155312, + 0.008144170977175236, + 0.0070936367847025394, + 0.012278646230697632, + 0.010652721859514713, + 0.024110594764351845, + 0.0238463394343853, + 0.01820378191769123, + 0.00933091901242733, + 0.011909820139408112, + 0.014427032321691513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0038330908864736557 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040984689258039 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.0040984689258039, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040984689258039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07946611195802689, + 0.04957138001918793, + 0.0535481721162796, + 0.08754514902830124, + 0.09478699415922165, + 0.06171109527349472, + 0.14626218378543854, + 0.07271268218755722, + 0.2100491225719452, + 0.14434708654880524 + ], + "credit_map": [ + 0.007103763986378908, + 0.007849566638469696, + 0.01761154644191265, + 0.007180641870945692, + 0.006040176842361689, + 0.0125197134912014, + 0.008838863112032413, + 0.012321840971708298, + 0.008644857443869114, + 0.004025641363114119, + 0.008508481085300446, + 0.004416614305227995, + 0.0036786883138120174, + 0.008852286264300346, + 0.006237014662474394, + 0.005207794718444347, + 0.004624651744961739, + 0.009110512211918831, + 0.005853165406733751, + 0.008538015186786652, + 0.009294385090470314, + 0.0058072651736438274, + 0.007397604640573263, + 0.0029225742910057306, + 0.012529278174042702, + 0.013392305932939053, + 0.0109457578510046, + 0.01276510488241911, + 0.013581186532974243, + 0.011651177890598774, + 0.007558549288660288, + 0.0051217894069850445, + 0.006644133012741804, + 0.008972152136266232, + 0.013806811533868313, + 0.00981627032160759, + 0.01022395584732294, + 0.017954058945178986, + 0.01579147018492222, + 0.011578142642974854, + 0.005343657452613115, + 0.005892250686883926, + 0.011618047952651978, + 0.0054596057161688805, + 0.007555045187473297, + 0.008133627474308014, + 0.009097513742744923, + 0.008611349388957024, + 0.01046580821275711, + 0.007585473824292421, + 0.014204266481101513, + 0.011009669862687588, + 0.04075516760349274, + 0.035547707229852676, + 0.011357136070728302, + 0.015336943790316582, + 0.012078669853508472, + 0.007223067339509726, + 0.011026738211512566, + 0.00893352646380663, + 0.007228120695799589, + 0.012730443850159645, + 0.007173723075538874, + 0.006318394560366869, + 0.0240741316229105, + 0.03792097046971321, + 0.028929945081472397, + 0.021666958928108215, + 0.03501390293240547, + 0.01616656966507435, + 0.021099796518683434, + 0.02517685294151306, + 0.01722966693341732, + 0.014083337038755417, + 0.022171158343553543, + 0.019652796909213066, + 0.02107001096010208, + 0.011773955076932907, + 0.014343065209686756, + 0.02402309700846672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0040984689258039 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977963674813509 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.00977963674813509, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977963674813509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12108750641345978, + 0.08166994154453278, + 0.2611347436904907, + 0.09373245388269424, + 0.05904430150985718, + 0.033656779676675797, + 0.054981913417577744, + 0.06211189925670624, + 0.09903988242149353, + 0.13354060053825378 + ], + "credit_map": [ + 0.012844551354646683, + 0.01145351305603981, + 0.02348308451473713, + 0.009123553521931171, + 0.01107631903141737, + 0.025584569200873375, + 0.014471913687884808, + 0.013050002045929432, + 0.007107369136065245, + 0.0067681134678423405, + 0.015854611992836, + 0.005788048729300499, + 0.007677752990275621, + 0.018016532063484192, + 0.008122689090669155, + 0.012334817089140415, + 0.02441803365945816, + 0.0410078801214695, + 0.03522047773003578, + 0.028308862820267677, + 0.04822670295834541, + 0.0398053415119648, + 0.021944483742117882, + 0.022202953696250916, + 0.016989691182971, + 0.008565799333155155, + 0.009352338500320911, + 0.006510334089398384, + 0.018904905766248703, + 0.022035900503396988, + 0.0057629807852208614, + 0.0056105004623532295, + 0.0053901528008282185, + 0.009119498543441296, + 0.00770421652123332, + 0.007143236696720123, + 0.008665360510349274, + 0.00887532439082861, + 0.006660550367087126, + 0.005485960282385349, + 0.003317232010886073, + 0.0046342527493834496, + 0.005415796767920256, + 0.0035879730712622404, + 0.004182782024145126, + 0.005266081541776657, + 0.003396473126485944, + 0.003856187453493476, + 0.005978515837341547, + 0.004194287117570639, + 0.00602089473977685, + 0.00379402213729918, + 0.019962510094046593, + 0.008051649667322636, + 0.0040209414437413216, + 0.002959091681987047, + 0.0102731017395854, + 0.00541763985529542, + 0.009189657866954803, + 0.0064755515195429325, + 0.009097209200263023, + 0.005654201377183199, + 0.008705442771315575, + 0.007299092132598162, + 0.014939445070922375, + 0.012857477180659771, + 0.013455796055495739, + 0.011087562888860703, + 0.012862621806561947, + 0.009420688264071941, + 0.010671486146748066, + 0.013744802214205265, + 0.014961213804781437, + 0.01609424687922001, + 0.0172678604722023, + 0.021015726029872894, + 0.01919778622686863, + 0.013192491605877876, + 0.016342954710125923, + 0.015468318946659565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.00977963674813509 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12280017137527466, + 0.17286734282970428, + 0.24881942570209503, + 0.03490216284990311, + 0.03266550227999687, + 0.05174342542886734, + 0.036869268864393234, + 0.03401989862322807, + 0.07798871397972107, + 0.18732404708862305 + ], + "credit_map": [ + 0.013543668203055859, + 0.014001631177961826, + 0.027311382815241814, + 0.011328272521495819, + 0.012056756764650345, + 0.02119854837656021, + 0.011597944423556328, + 0.011761960573494434, + 0.019363779574632645, + 0.01264205016195774, + 0.03570608049631119, + 0.013082385063171387, + 0.011707662604749203, + 0.034483179450035095, + 0.016251204535365105, + 0.0296309906989336, + 0.04042370989918709, + 0.03573158010840416, + 0.030477086082100868, + 0.019811786711215973, + 0.04287341237068176, + 0.04194135218858719, + 0.02005128748714924, + 0.017509212717413902, + 0.004903580993413925, + 0.003776158671826124, + 0.006958402693271637, + 0.002651310060173273, + 0.0032062388490885496, + 0.006983723491430283, + 0.0029862120281904936, + 0.0034365353640168905, + 0.0026952610351145267, + 0.005492647178471088, + 0.00418435363098979, + 0.003389971563592553, + 0.00323619251139462, + 0.005483538843691349, + 0.0037506981752812862, + 0.004432839807122946, + 0.00432360777631402, + 0.005739920772612095, + 0.010828442871570587, + 0.005158240906894207, + 0.005244531203061342, + 0.010874585248529911, + 0.0039634546265006065, + 0.005610642954707146, + 0.00404925225302577, + 0.00368672888725996, + 0.005903129931539297, + 0.004226056393235922, + 0.008135410025715828, + 0.004023612476885319, + 0.0031435522250831127, + 0.0037015276029706, + 0.003900150302797556, + 0.0025891531258821487, + 0.004536914173513651, + 0.003716510022059083, + 0.004378543235361576, + 0.007529315538704395, + 0.004042870365083218, + 0.0033264420926570892, + 0.009150301106274128, + 0.007700638845562935, + 0.01879229210317135, + 0.007159579545259476, + 0.010283974930644035, + 0.01158933062106371, + 0.006781792268157005, + 0.006530805956572294, + 0.019159045070409775, + 0.016590651124715805, + 0.028273390606045723, + 0.035857412964105606, + 0.027444979175925255, + 0.02236093021929264, + 0.01894667185842991, + 0.018690964207053185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606426201760769 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606426201760769, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606426201760769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.146536186337471, + 0.054816536605358124, + 0.09865725040435791, + 0.07225895673036575, + 0.04073265194892883, + 0.05545811355113983, + 0.09773995727300644, + 0.0691443383693695, + 0.22366520762443542, + 0.14099077880382538 + ], + "credit_map": [ + 0.01622879132628441, + 0.021096156910061836, + 0.03006168268620968, + 0.01241588220000267, + 0.0158548541367054, + 0.01988762803375721, + 0.015161402523517609, + 0.015829795971512794, + 0.006975538097321987, + 0.0044230199418962, + 0.01092544849961996, + 0.004263250157237053, + 0.003676760010421276, + 0.007719358429312706, + 0.007095725741237402, + 0.00973743386566639, + 0.012396949343383312, + 0.014780248515307903, + 0.012844763696193695, + 0.008046921342611313, + 0.01836639828979969, + 0.01745556853711605, + 0.007241084240376949, + 0.007525317836552858, + 0.006290747784078121, + 0.00900556892156601, + 0.014166349545121193, + 0.009947258979082108, + 0.006516103632748127, + 0.017195317894220352, + 0.005595613736659288, + 0.0035420015919953585, + 0.004221755545586348, + 0.006060811225324869, + 0.006102775689214468, + 0.005369036458432674, + 0.0038570105098187923, + 0.006445840932428837, + 0.004827255383133888, + 0.003848171327263117, + 0.005195081699639559, + 0.008565106429159641, + 0.009759604930877686, + 0.004115001298487186, + 0.009657067246735096, + 0.005926823243498802, + 0.006419738754630089, + 0.005819688085466623, + 0.010424449108541012, + 0.007913501933217049, + 0.010635613463819027, + 0.010661202482879162, + 0.015664923936128616, + 0.020816879346966743, + 0.009344003163278103, + 0.012279387563467026, + 0.005348629783838987, + 0.004972216673195362, + 0.01920877955853939, + 0.007148818112909794, + 0.007457949221134186, + 0.014861708506941795, + 0.005928778555244207, + 0.004217465408146381, + 0.02555500715970993, + 0.03677339479327202, + 0.0391223281621933, + 0.02375713549554348, + 0.034189458936452866, + 0.01774870604276657, + 0.021456385031342506, + 0.025062792003154755, + 0.014467972330749035, + 0.014171759597957134, + 0.02462751418352127, + 0.028895847499370575, + 0.021087557077407837, + 0.011628820560872555, + 0.012934288941323757, + 0.013177015818655491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.01606426201760769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, swooping shot begins at a gentle, high angle, then slowly descends and advances, gently following the movements of a delicate pansy. The flower is in full bloom, its velvety purple and yellow petals, embellished with delicate dark veins, standing out against a carpet of thick, emerald green, and moist moss. A soft, warm yellow light streams down from above and sideways, penetrating the edges of the petals, giving them a translucent quality and casting a dappled, warm glow across the moss. Under the lens's gaze, the pansy's stem delicately descends into the soft moss with an almost imperceptible, graceful rhythm. The green moss fibers gently part, welcoming the flower, like the meticulous construction of a cozy nest. A shallow depth of field gently blurs the background, focusing all attention on this microcosmic ritual of life, creating an atmosphere of profound tranquility, healing, and vibrant warmth.", + "seed": 2084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_1070_seed_2084_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": -0.009200215665623545, + "vlm_logprob": -0.009200215665623545, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.009200215665623545, + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004340988118201494 + ], + "opposite_logprob_tokens": [ + -5.879341125488281 + ], + "logprob_reward": -0.004340988118201494, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004340988118201494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12052151560783386, + 0.05930156260728836, + 0.1733868569135666, + 0.05352048575878143, + 0.10232532769441605, + 0.0442328155040741, + 0.12450484931468964, + 0.03138957545161247, + 0.07287853211164474, + 0.21793852746486664 + ], + "credit_map": [ + 0.024683360010385513, + 0.010222522541880608, + 0.009785539470613003, + 0.008660104125738144, + 0.007680579554289579, + 0.028849484398961067, + 0.015390110202133656, + 0.015249812044203281, + 0.013545132242143154, + 0.00393840903416276, + 0.006505158729851246, + 0.004578020423650742, + 0.004098421894013882, + 0.004807289224117994, + 0.007730779703706503, + 0.014098353683948517, + 0.022249216213822365, + 0.014835306443274021, + 0.04812869429588318, + 0.02365141734480858, + 0.012034053914248943, + 0.013940846547484398, + 0.015716826543211937, + 0.02283048816025257, + 0.007798286620527506, + 0.0027628829702734947, + 0.015874015167355537, + 0.005689980462193489, + 0.0022651113104075193, + 0.0045728739351034164, + 0.004869733937084675, + 0.009687598794698715, + 0.015456093475222588, + 0.006979559548199177, + 0.03042963333427906, + 0.009568827226758003, + 0.007004577666521072, + 0.007310896180570126, + 0.0068021537736058235, + 0.018773583695292473, + 0.0032977100927382708, + 0.00369782792404294, + 0.008015808649361134, + 0.0033146110363304615, + 0.004039521794766188, + 0.012421674095094204, + 0.002861810615286231, + 0.006583855487406254, + 0.007079951465129852, + 0.006904374808073044, + 0.010675578378140926, + 0.010489272885024548, + 0.010811707004904747, + 0.031151991337537766, + 0.035190399736166, + 0.01220157090574503, + 0.0033578502479940653, + 0.002960923593491316, + 0.007249548565596342, + 0.0035936504136770964, + 0.0035285146441310644, + 0.004493725020438433, + 0.0028029114473611116, + 0.0034024512860924006, + 0.008586016483604908, + 0.006044498644769192, + 0.013445712625980377, + 0.021564381197094917, + 0.0034319069236516953, + 0.005581021308898926, + 0.007953275926411152, + 0.0062717231921851635, + 0.012180179357528687, + 0.008247248828411102, + 0.08289256691932678, + 0.07792283594608307, + 0.0037942358758300543, + 0.01003579143434763, + 0.014613498002290726, + 0.008252155035734177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.004340988118201494 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028237728402018547 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.0028237728402018547, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028237728402018547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0357804000377655, + 0.01841985248029232, + 0.11194663494825363, + 0.3224233090877533, + 0.05061589181423187, + 0.14621208608150482, + 0.06652924418449402, + 0.044912125915288925, + 0.12082456052303314, + 0.08233585953712463 + ], + "credit_map": [ + 0.005193488672375679, + 0.0035552321933209896, + 0.004396494012326002, + 0.004109502304345369, + 0.002630089409649372, + 0.0065459152683615685, + 0.003414884675294161, + 0.0059347981587052345, + 0.004372075665742159, + 0.001183116459287703, + 0.005219228100031614, + 0.0014560596318915486, + 0.0008724585059098899, + 0.0018035672837868333, + 0.0014374175807461143, + 0.002075932687148452, + 0.01359191071242094, + 0.010301140137016773, + 0.030230173841118813, + 0.012839001603424549, + 0.009458030574023724, + 0.008986657485365868, + 0.011892233975231647, + 0.014647496864199638, + 0.052013374865055084, + 0.010688875801861286, + 0.13551707565784454, + 0.01796211116015911, + 0.005338207818567753, + 0.017007572576403618, + 0.011734498664736748, + 0.07216163724660873, + 0.007586443796753883, + 0.004585725720971823, + 0.011546826921403408, + 0.00637864600867033, + 0.004216555040329695, + 0.004820558708161116, + 0.0038645160384476185, + 0.0076166256330907345, + 0.013391599990427494, + 0.006545309443026781, + 0.04654113948345184, + 0.019883589819073677, + 0.004536676220595837, + 0.031182222068309784, + 0.009253288619220257, + 0.014878282323479652, + 0.0038162635173648596, + 0.006413428578525782, + 0.006511325016617775, + 0.007281138561666012, + 0.007145250681787729, + 0.011409848928451538, + 0.017270822077989578, + 0.006681167520582676, + 0.004389883019030094, + 0.006031667347997427, + 0.0068410784006118774, + 0.006055898033082485, + 0.0057101319544017315, + 0.006006474606692791, + 0.005567539948970079, + 0.004309457261115313, + 0.012043379247188568, + 0.014912073500454426, + 0.019135959446430206, + 0.02053207904100418, + 0.009554092772305012, + 0.017033766955137253, + 0.01267533004283905, + 0.01493788417428732, + 0.00745585048571229, + 0.007286482490599155, + 0.014059053733944893, + 0.023551268503069878, + 0.003548659151419997, + 0.005961098708212376, + 0.006877831649035215, + 0.01359561923891306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.0028237728402018547 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577285438776016 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577285438776016, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577285438776016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04916945472359657, + 0.02435987815260887, + 0.0558512918651104, + 0.3394618034362793, + 0.09757846593856812, + 0.036020103842020035, + 0.036695968359708786, + 0.03068321943283081, + 0.0738215446472168, + 0.2563582956790924 + ], + "credit_map": [ + 0.007850910536944866, + 0.0038057328201830387, + 0.005823342129588127, + 0.004446546547114849, + 0.0036553966347128153, + 0.008527219295501709, + 0.005807101726531982, + 0.009253204800188541, + 0.006870895158499479, + 0.0015869095223024487, + 0.003932461608201265, + 0.001459510182030499, + 0.0008785475511103868, + 0.0017264618072658777, + 0.0026319585740566254, + 0.005273131188005209, + 0.006202602293342352, + 0.006830720696598291, + 0.012277577072381973, + 0.005901670549064875, + 0.005773226730525494, + 0.004440529737621546, + 0.007157318759709597, + 0.0072676390409469604, + 0.05221742019057274, + 0.010848533362150192, + 0.14173385500907898, + 0.021748535335063934, + 0.006710494868457317, + 0.017728401347994804, + 0.01208555605262518, + 0.07638899236917496, + 0.011229299940168858, + 0.011376786045730114, + 0.015658752992749214, + 0.016907287761569023, + 0.01060805469751358, + 0.00980279128998518, + 0.009687742218375206, + 0.012307754717767239, + 0.003010410815477371, + 0.0012011409271508455, + 0.012027458287775517, + 0.0029590362682938576, + 0.0020626718178391457, + 0.008685929700732231, + 0.0027392073534429073, + 0.0033342468086630106, + 0.0020505287684500217, + 0.00310824834741652, + 0.003623856930062175, + 0.003288004081696272, + 0.004381038248538971, + 0.007240057922899723, + 0.010002750903367996, + 0.0030014794319868088, + 0.0028967263642698526, + 0.003201895160600543, + 0.005873518995940685, + 0.0035603800788521767, + 0.0025688803289085627, + 0.0062728943303227425, + 0.0029869615100324154, + 0.0033219598699361086, + 0.00877009704709053, + 0.0070173912681639194, + 0.011145409196615219, + 0.02166949212551117, + 0.0034413968678563833, + 0.006574308034032583, + 0.007676344830542803, + 0.0075270989909768105, + 0.01429841760545969, + 0.011864569038152695, + 0.10007720440626144, + 0.08754142373800278, + 0.004556052852421999, + 0.011900201439857483, + 0.01551028061658144, + 0.010610118508338928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.020577285438776016 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021972341928631067 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.0021972341928631067, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021972341928631067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07236422598361969, + 0.03243415802717209, + 0.1303340196609497, + 0.05993978679180145, + 0.09766891598701477, + 0.2534206807613373, + 0.12135403603315353, + 0.0702928826212883, + 0.05307019501924515, + 0.10912109911441803 + ], + "credit_map": [ + 0.008945010602474213, + 0.004858973901718855, + 0.007884757593274117, + 0.006041655782610178, + 0.003983344882726669, + 0.015068349428474903, + 0.006554687861353159, + 0.019027452915906906, + 0.004597931168973446, + 0.002817666856572032, + 0.006258430890738964, + 0.0026755500584840775, + 0.0031518004834651947, + 0.0037525889929383993, + 0.004004696849733591, + 0.005175494588911533, + 0.018986253067851067, + 0.009823904372751713, + 0.03107239119708538, + 0.01679600030183792, + 0.00904335081577301, + 0.01024686824530363, + 0.01321693416684866, + 0.0211483146995306, + 0.007462238892912865, + 0.003322411561384797, + 0.016744107007980347, + 0.00739542581140995, + 0.0043274289928376675, + 0.003905915655195713, + 0.007170758675783873, + 0.009611501358449459, + 0.014375471509993076, + 0.008946290239691734, + 0.025795983150601387, + 0.010755359195172787, + 0.006744179408997297, + 0.007680491544306278, + 0.007393790874630213, + 0.015977365896105766, + 0.02039068564772606, + 0.00612293416634202, + 0.101931631565094, + 0.018229510635137558, + 0.006914445664733648, + 0.06214311346411705, + 0.009279555641114712, + 0.028408844023942947, + 0.005955457221716642, + 0.00837732944637537, + 0.0077933212742209435, + 0.009135291911661625, + 0.010768655687570572, + 0.02891804277896881, + 0.038584720343351364, + 0.011821226216852665, + 0.0064344704151153564, + 0.008307287469506264, + 0.011824297718703747, + 0.008052535355091095, + 0.008636262267827988, + 0.013143242336809635, + 0.00649629021063447, + 0.007398501504212618, + 0.007639268413186073, + 0.006294667720794678, + 0.005933638662099838, + 0.014192956499755383, + 0.004192053340375423, + 0.004637884441763163, + 0.005701558198779821, + 0.004478174727410078, + 0.006881400942802429, + 0.0051337736658751965, + 0.03863343596458435, + 0.03193257004022598, + 0.003918100148439407, + 0.006400196813046932, + 0.00787228625267744, + 0.008349350653588772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.0021972341928631067 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061797738075256 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061797738075256, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061797738075256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053332723677158356, + 0.02609783224761486, + 0.03896189481019974, + 0.32362672686576843, + 0.01767687499523163, + 0.10359779745340347, + 0.128741055727005, + 0.04635661095380783, + 0.05284423008561134, + 0.20876432955265045 + ], + "credit_map": [ + 0.015889916568994522, + 0.003378927009180188, + 0.004199439659714699, + 0.003016537055373192, + 0.002279294189065695, + 0.01178281381726265, + 0.003983302041888237, + 0.008802485652267933, + 0.0035117820370942354, + 0.002658227691426873, + 0.0029171379283070564, + 0.0024772752076387405, + 0.0019200901733711362, + 0.002208153484389186, + 0.004672612063586712, + 0.005732550285756588, + 0.0049389130435884, + 0.00280874315649271, + 0.012137090787291527, + 0.004971490241587162, + 0.0030332652386277914, + 0.0031993919983506203, + 0.0029133684001863003, + 0.004959627520292997, + 0.052192673087120056, + 0.01114702969789505, + 0.13564591109752655, + 0.016359858214855194, + 0.005891456268727779, + 0.01761581189930439, + 0.01198993157595396, + 0.07278401404619217, + 0.0023132849019020796, + 0.00160212442278862, + 0.00235741026699543, + 0.0028143124654889107, + 0.0012913693208247423, + 0.002197603927925229, + 0.00248096720315516, + 0.002619800390675664, + 0.00910877250134945, + 0.0034253865014761686, + 0.03797207027673721, + 0.008985153399407864, + 0.006298556458204985, + 0.01760614477097988, + 0.004764389246702194, + 0.015437317080795765, + 0.005997328087687492, + 0.00898693222552538, + 0.009217395447194576, + 0.010827179066836834, + 0.010808986611664295, + 0.03271275758743286, + 0.03807293623685837, + 0.012117523699998856, + 0.004745988175272942, + 0.006716795265674591, + 0.00726556358858943, + 0.005133684724569321, + 0.006785696372389793, + 0.004982187878340483, + 0.005243366118520498, + 0.0054833246394991875, + 0.004145370330661535, + 0.005531562026590109, + 0.012003935873508453, + 0.011091399937868118, + 0.0042183492332696915, + 0.006034885998815298, + 0.004680770915001631, + 0.005137950647622347, + 0.01189460139721632, + 0.009268201887607574, + 0.0708647221326828, + 0.07841361314058304, + 0.005414700601249933, + 0.01133058499544859, + 0.015928838402032852, + 0.005649040453135967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.016061797738075256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, swooping shot begins at a gentle, high angle, then slowly descends and advances, gently following the movements of a delicate pansy. The flower is in full bloom, its velvety purple and yellow petals, embellished with delicate dark veins, standing out against a carpet of thick, emerald green, and moist moss. A soft, warm yellow light streams down from above and sideways, penetrating the edges of the petals, giving them a translucent quality and casting a dappled, warm glow across the moss. Under the lens's gaze, the pansy's stem delicately descends into the soft moss with an almost imperceptible, graceful rhythm. The green moss fibers gently part, welcoming the flower, like the meticulous construction of a cozy nest. A shallow depth of field gently blurs the background, focusing all attention on this microcosmic ritual of life, creating an atmosphere of profound tranquility, healing, and vibrant warmth.", + "seed": 2085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_8/index_1070_seed_2085_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": -0.009340270422399044, + "vlm_logprob": -0.009340270422399044, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.009340270422399044, + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006454454734921455 + ], + "opposite_logprob_tokens": [ + -5.1314544677734375 + ], + "logprob_reward": -0.006454454734921455, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006454454734921455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17846040427684784, + 0.063291035592556, + 0.07351335883140564, + 0.1051507294178009, + 0.04738494008779526, + 0.05977829545736313, + 0.37318155169487, + 0.022862855345010757, + 0.01347349677234888, + 0.06290334463119507 + ], + "credit_map": [ + 0.019713198766112328, + 0.023827413097023964, + 0.022301888093352318, + 0.02782292105257511, + 0.026480546221137047, + 0.02150150015950203, + 0.01519775390625, + 0.02161518670618534, + 0.01056582573801279, + 0.0141737787052989, + 0.007770557422190905, + 0.009140689857304096, + 0.00497661717236042, + 0.004043614026159048, + 0.006082622334361076, + 0.006537329871207476, + 0.016868717968463898, + 0.010905151255428791, + 0.009591108188033104, + 0.00837212335318327, + 0.004692667163908482, + 0.005209631286561489, + 0.006361552979797125, + 0.011512408033013344, + 0.008805271238088608, + 0.009742673486471176, + 0.01486016996204853, + 0.014071784913539886, + 0.009449083358049393, + 0.008651037700474262, + 0.02643599733710289, + 0.013134723529219627, + 0.004878629464656115, + 0.0053487252444028854, + 0.00755552900955081, + 0.006466641090810299, + 0.007894671522080898, + 0.006190691143274307, + 0.00479137571528554, + 0.0042586782947182655, + 0.005397444125264883, + 0.005541839171200991, + 0.009056023322045803, + 0.012232865206897259, + 0.007081883959472179, + 0.006131425965577364, + 0.00676715699955821, + 0.0075696599669754505, + 0.06060413271188736, + 0.11499913781881332, + 0.0224442258477211, + 0.024758540093898773, + 0.027492448687553406, + 0.07219017297029495, + 0.035193122923374176, + 0.015499792993068695, + 0.0028211278840899467, + 0.002905626315623522, + 0.002873420948162675, + 0.002910661045461893, + 0.003941578324884176, + 0.0028701687697321177, + 0.00228378945030272, + 0.002256485866382718, + 0.0012584167998284101, + 0.0020301842596381903, + 0.0018083908362314105, + 0.00150672800373286, + 0.002242720453068614, + 0.0017335190204903483, + 0.0015984541969373822, + 0.001295084715820849, + 0.005991077981889248, + 0.006157408468425274, + 0.01136940997093916, + 0.008578430861234665, + 0.0058028679341077805, + 0.009628531523048878, + 0.007653622888028622, + 0.007721998728811741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.006454454734921455 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.097022145986557, + 0.11515438556671143, + 0.16689527034759521, + 0.1931612491607666, + 0.05586475878953934, + 0.027393724769353867, + 0.1892753541469574, + 0.040168412029743195, + 0.03058192878961563, + 0.08448286354541779 + ], + "credit_map": [ + 0.010465241968631744, + 0.020788365975022316, + 0.015364515595138073, + 0.013207759708166122, + 0.00844703521579504, + 0.010688078589737415, + 0.005875678267329931, + 0.012185460887849331, + 0.014590241014957428, + 0.009718935005366802, + 0.014618702232837677, + 0.016154281795024872, + 0.011880656704306602, + 0.010421942919492722, + 0.02138170413672924, + 0.016387905925512314, + 0.03310180827975273, + 0.019190605729818344, + 0.020326348021626472, + 0.027755454182624817, + 0.012691128998994827, + 0.013179556466639042, + 0.017858309671282768, + 0.022792045027017593, + 0.023253094404935837, + 0.016735000535845757, + 0.02773224003612995, + 0.032500069588422775, + 0.013010850176215172, + 0.01377796195447445, + 0.039862312376499176, + 0.026289690285921097, + 0.005178486462682486, + 0.00832423847168684, + 0.0051696947775781155, + 0.006593299098312855, + 0.01356545276939869, + 0.0059297773987054825, + 0.00636414997279644, + 0.004739656578749418, + 0.0024230722337961197, + 0.0034025816712528467, + 0.0036784051917493343, + 0.005533489864319563, + 0.0041898200288414955, + 0.003208319889381528, + 0.002544279443100095, + 0.002413753652945161, + 0.020232610404491425, + 0.019084351137280464, + 0.015530572272837162, + 0.01919875480234623, + 0.014734582044184208, + 0.02162306196987629, + 0.058339446783065796, + 0.020531944930553436, + 0.0033640735782682896, + 0.004129169508814812, + 0.009935039095580578, + 0.005033369176089764, + 0.0044168587774038315, + 0.004068492446094751, + 0.004821069072932005, + 0.004400336649268866, + 0.0033030277118086815, + 0.004142400808632374, + 0.00516614830121398, + 0.0037756040692329407, + 0.003109476063400507, + 0.00281791971065104, + 0.0036114188842475414, + 0.004655931610614061, + 0.008769412524998188, + 0.00985078327357769, + 0.01418209820985794, + 0.008824935182929039, + 0.008623571135103703, + 0.017934534698724747, + 0.007740369997918606, + 0.00855715200304985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.00525747099891305 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06423648446798325, + 0.5496116280555725, + 0.09208674728870392, + 0.04093454033136368, + 0.0231378935277462, + 0.04537251219153404, + 0.10314483195543289, + 0.03943353518843651, + 0.00907083973288536, + 0.03297095000743866 + ], + "credit_map": [ + 0.00627975445240736, + 0.013381178490817547, + 0.012749104760587215, + 0.007644259370863438, + 0.006043646484613419, + 0.006592616904526949, + 0.005815768614411354, + 0.005730160046368837, + 0.06976839900016785, + 0.11442077159881592, + 0.07502292841672897, + 0.10139230638742447, + 0.042703352868556976, + 0.030573807656764984, + 0.052658937871456146, + 0.06307114660739899, + 0.01909743808209896, + 0.010486911050975323, + 0.01275477185845375, + 0.012146768160164356, + 0.006701227277517319, + 0.007359097246080637, + 0.010607410222291946, + 0.012933124788105488, + 0.004375016782432795, + 0.005988320801407099, + 0.004162339027971029, + 0.004919090308248997, + 0.0026035839691758156, + 0.002880505286157131, + 0.009846665896475315, + 0.00615901593118906, + 0.0020029048901051283, + 0.002771241357550025, + 0.0021565728820860386, + 0.005241028033196926, + 0.003288585925474763, + 0.0022120175417512655, + 0.003593375673517585, + 0.001872168155387044, + 0.0037238397635519505, + 0.006191552150994539, + 0.009159703738987446, + 0.006689512170851231, + 0.007913192734122276, + 0.0040807886980473995, + 0.0037962179630994797, + 0.0038177098613232374, + 0.008544874377548695, + 0.019902445375919342, + 0.006768187042325735, + 0.009857263416051865, + 0.010163795202970505, + 0.00833979807794094, + 0.029236696660518646, + 0.010331772267818451, + 0.0031192167662084103, + 0.004208816215395927, + 0.007779960986226797, + 0.007196536287665367, + 0.005317370872944593, + 0.003382396651431918, + 0.004687961656600237, + 0.0037412738893181086, + 0.000836805731523782, + 0.001260683056898415, + 0.001460629515349865, + 0.0013216458028182387, + 0.0007843036437407136, + 0.0007960686343722045, + 0.0011023475090041757, + 0.001508355955593288, + 0.003222613362595439, + 0.0036829630844295025, + 0.005642523989081383, + 0.003952506463974714, + 0.004088066518306732, + 0.00596720352768898, + 0.0034791347570717335, + 0.0029359383042901754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.009779755026102066 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635422024875879 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635422024875879, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635422024875879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12853263318538666, + 0.18066450953483582, + 0.11995655298233032, + 0.05378532037138939, + 0.05182361975312233, + 0.05737665668129921, + 0.27779605984687805, + 0.03251461312174797, + 0.039111603051424026, + 0.058438491076231 + ], + "credit_map": [ + 0.01097738929092884, + 0.018931932747364044, + 0.01958598755300045, + 0.01786850579082966, + 0.019572563469409943, + 0.015349839814007282, + 0.01215688418596983, + 0.01408953033387661, + 0.020197652280330658, + 0.03404980152845383, + 0.02892357110977173, + 0.02612914703786373, + 0.019187230616807938, + 0.013537739403545856, + 0.017448239028453827, + 0.02119113691151142, + 0.027430308982729912, + 0.011239676736295223, + 0.019119998440146446, + 0.013529131188988686, + 0.008129996247589588, + 0.009021727368235588, + 0.01715158298611641, + 0.014334128238260746, + 0.006350503768771887, + 0.008416717872023582, + 0.005839848425239325, + 0.006297818385064602, + 0.0059737833216786385, + 0.005243345629423857, + 0.00820972677320242, + 0.00745357433333993, + 0.005435994826257229, + 0.006194411776959896, + 0.00541047053411603, + 0.007340034935623407, + 0.010943405330181122, + 0.007555882912129164, + 0.0040357839316129684, + 0.004907635506242514, + 0.0053244829177856445, + 0.0050300294533371925, + 0.012738378718495369, + 0.007431883364915848, + 0.007142505142837763, + 0.00788182858377695, + 0.004663005471229553, + 0.007164540234953165, + 0.026173558086156845, + 0.05523402988910675, + 0.020073574036359787, + 0.03312741965055466, + 0.02570992335677147, + 0.013754168525338173, + 0.07563553005456924, + 0.028087858110666275, + 0.003336035879328847, + 0.003146551549434662, + 0.007328754756599665, + 0.004362017381936312, + 0.003613016102463007, + 0.004262102302163839, + 0.003701217006891966, + 0.0027649165131151676, + 0.0037109144032001495, + 0.00628657266497612, + 0.006127069238573313, + 0.0038065840490162373, + 0.0042138779535889626, + 0.004416697658598423, + 0.0048836832866072655, + 0.005666201002895832, + 0.00570765882730484, + 0.006513711530715227, + 0.009776689112186432, + 0.01055830530822277, + 0.005955742672085762, + 0.006763400509953499, + 0.007178924977779388, + 0.005984053947031498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004635422024875879 + }, + { + "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 delicate pansy flower in full bloom with velvety purple and yellow petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pansy standing on a stem that descends into a carpet of thick, emerald green, moist moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by soft, warm yellow light streaming from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a shallow depth of field that blurs the background to focus on the flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin at a gentle high angle and slowly descend while advancing to follow the flower?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057424932718277 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057424932718277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057424932718277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11132800579071045, + 0.07936873286962509, + 0.25890621542930603, + 0.08978480845689774, + 0.09843457490205765, + 0.043480146676301956, + 0.24105720221996307, + 0.012161328457295895, + 0.01967635191977024, + 0.0458025224506855 + ], + "credit_map": [ + 0.010587203316390514, + 0.020620914176106453, + 0.021231403574347496, + 0.015986349433660507, + 0.011372826993465424, + 0.008622978813946247, + 0.009110321290791035, + 0.013796010985970497, + 0.009849135763943195, + 0.009669297374784946, + 0.00859108380973339, + 0.010503648780286312, + 0.008289511315524578, + 0.006451721768826246, + 0.01065223477780819, + 0.015362103469669819, + 0.05206181854009628, + 0.030026348307728767, + 0.03451938554644585, + 0.03567192703485489, + 0.02028629370033741, + 0.02197793871164322, + 0.028141507878899574, + 0.03622099757194519, + 0.011376535519957542, + 0.01248590275645256, + 0.013230415992438793, + 0.014631258323788643, + 0.009321548044681549, + 0.006736369803547859, + 0.012875991873443127, + 0.009126780554652214, + 0.008742660284042358, + 0.011522178538143635, + 0.012950228527188301, + 0.015002232044935226, + 0.016427211463451385, + 0.015332356095314026, + 0.010095388628542423, + 0.008362320251762867, + 0.004333794116973877, + 0.005856161937117577, + 0.01259702444076538, + 0.005152109544724226, + 0.004296233877539635, + 0.0036214706487953663, + 0.004011794924736023, + 0.00361155834980309, + 0.03777386620640755, + 0.09001962095499039, + 0.010861788876354694, + 0.015132133848965168, + 0.02213507890701294, + 0.039860695600509644, + 0.014553399756550789, + 0.010720609687268734, + 0.0012781019322574139, + 0.0016852379776537418, + 0.0016286653699353337, + 0.001875832094810903, + 0.0018075350672006607, + 0.0014140605926513672, + 0.0011793082812801003, + 0.0012925869086757302, + 0.002049192786216736, + 0.003013444133102894, + 0.0027759550139307976, + 0.0031373833771795034, + 0.0023125573061406612, + 0.002552354009822011, + 0.0021085029002279043, + 0.0017269618110731244, + 0.004642230924218893, + 0.005394334904849529, + 0.006543201860040426, + 0.005419926717877388, + 0.004982453770935535, + 0.010173500515520573, + 0.0039002755656838417, + 0.004746597725898027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.02057424932718277 + } + ], + "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/0000167/group_9/index_tensor([9428, 9428])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_tensor([9428, 9428])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eed22fa6a5e87ab2992d4c2af982abb2a57b6b8e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_tensor([9428, 9428])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of a sophisticated industrial product photography, a perfectly fixed macro close-up steadily captures the entire dynamic process. The background is an absolute, non-reflective black void, directing all visual focus to the central subject. Under bright, even commercial lighting, a round pressed powder sits motionless in the center of the frame. Its surface, a velvety matte, is embossed with a nebula-like pattern of saturated colors: vibrant magenta, deep cyan, and brilliant lemon yellow. A drop of crystal-clear, highly viscous liquid, like a perfect miniature glass sphere, slowly and precisely drips from the top of the frame. The moment the drop strikes the center of the powder, it instantly disrupts the powder's static equilibrium, triggering a miniature crown-shaped splash. The delicate, multicolored powders are then swept into the air, forming a dazzling and complex microscopic vortex. Driven by the liquid, these brilliant colors chase, swirl, stir, and merge with each other, like a silent color storm erupting on the millimeter scale. The whole process is full of scientific precision and stunning visual beauty, stably and clearly showing the dynamic interaction between materials.", + "seed": 9525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_9428_seed_9525_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": -0.16946503520011902, + "vlm_logprob": -0.16946503520011902, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.16946503520011902, + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.693987250328064 + ], + "opposite_logprob_tokens": [ + -0.693987250328064 + ], + "logprob_reward": -0.693987250328064, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693987250328064, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06234237924218178, + 0.03089681640267372, + 0.030818933621048927, + 0.049554068595170975, + 0.5508856773376465, + 0.1691388189792633, + 0.03134338930249214, + 0.05560268834233284, + 0.011436420492827892, + 0.00798072386533022 + ], + "credit_map": [ + 0.011726276949048042, + 0.0034136974718421698, + 0.0160903912037611, + 0.004196758382022381, + 0.0037399118300527334, + 0.008910986594855785, + 0.004538369830697775, + 0.009725991636514664, + 0.0056023732759058475, + 0.0015219530323520303, + 0.004166540689766407, + 0.0028795385733246803, + 0.0023551504127681255, + 0.004098802804946899, + 0.002371304901316762, + 0.007901154458522797, + 0.006266417913138866, + 0.001438584178686142, + 0.005647278856486082, + 0.0026520758401602507, + 0.0017237627180293202, + 0.004387753549963236, + 0.002123857382684946, + 0.006579206325113773, + 0.008229700848460197, + 0.004395580384880304, + 0.004331313073635101, + 0.004845615942031145, + 0.00721326470375061, + 0.010467583313584328, + 0.0035273621324449778, + 0.00654365262016654, + 0.1370425671339035, + 0.030270522460341454, + 0.03706962242722511, + 0.02802698127925396, + 0.03339682146906853, + 0.18600787222385406, + 0.025630494579672813, + 0.07344086468219757, + 0.03712642937898636, + 0.011351517401635647, + 0.02667488344013691, + 0.007998600602149963, + 0.011832586489617825, + 0.024422504007816315, + 0.009579970501363277, + 0.04015232250094414, + 0.004185480065643787, + 0.0016757883131504059, + 0.006202303804457188, + 0.0020360304042696953, + 0.0028620390221476555, + 0.003492100164294243, + 0.0022877336014062166, + 0.008601916953921318, + 0.02043670415878296, + 0.0020447152201086283, + 0.015667129307985306, + 0.0026918561197817326, + 0.0015538650332018733, + 0.004362392704933882, + 0.0018713120371103287, + 0.0069747194647789, + 0.0018135304562747478, + 0.0013380993623286486, + 0.0011442379327490926, + 0.0016765467589721084, + 0.0010928029660135508, + 0.0023984964936971664, + 0.0009177703759633005, + 0.0010549372527748346, + 0.001101642963476479, + 0.0008597149862907827, + 0.0006881011067889631, + 0.0008491415646858513, + 0.0020324448123574257, + 0.0011308030225336552, + 0.0007513240561820567, + 0.0005675516440533102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.693987250328064 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899230718612671 + ], + "opposite_logprob_tokens": [ + -2.5789923667907715 + ], + "logprob_reward": -0.07899230718612671, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899230718612671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09961856156587601, + 0.08696185052394867, + 0.04487992823123932, + 0.2528345286846161, + 0.19311869144439697, + 0.10191800445318222, + 0.06222926825284958, + 0.08479608595371246, + 0.04178575426340103, + 0.03185736760497093 + ], + "credit_map": [ + 0.017035773023962975, + 0.0064050364308059216, + 0.014175387099385262, + 0.010586260817945004, + 0.006476160138845444, + 0.010768985375761986, + 0.008543133735656738, + 0.025627827271819115, + 0.015325616113841534, + 0.0036961142905056477, + 0.010658588260412216, + 0.004922753665596247, + 0.005007995758205652, + 0.013414347544312477, + 0.007248847279697657, + 0.02668757550418377, + 0.008581605739891529, + 0.0025639452505856752, + 0.005337386392056942, + 0.002663484076038003, + 0.0027827159501612186, + 0.006520280614495277, + 0.0031158607453107834, + 0.013314646668732166, + 0.04886907339096069, + 0.027561111375689507, + 0.018482713028788567, + 0.01753082498908043, + 0.030709482729434967, + 0.06578831374645233, + 0.014806948602199554, + 0.029086029157042503, + 0.04339468479156494, + 0.012511900626122952, + 0.01608184538781643, + 0.014506648294627666, + 0.011980797164142132, + 0.05624514818191528, + 0.012961125932633877, + 0.02543654665350914, + 0.019244801253080368, + 0.011971301399171352, + 0.015543773770332336, + 0.008282962255179882, + 0.009803609922528267, + 0.012631717137992382, + 0.007868686690926552, + 0.016571154817938805, + 0.014179721474647522, + 0.004105753730982542, + 0.007011694833636284, + 0.005985944997519255, + 0.004219230730086565, + 0.008701645769178867, + 0.004995458293706179, + 0.01302981749176979, + 0.014934507198631763, + 0.004201650153845549, + 0.024632569402456284, + 0.0062734163366258144, + 0.006402248982340097, + 0.011722649447619915, + 0.005476347636431456, + 0.01115269586443901, + 0.005498726852238178, + 0.003923929296433926, + 0.004486030898988247, + 0.009984010830521584, + 0.003014881629496813, + 0.006072869058698416, + 0.005753226112574339, + 0.0030520788859575987, + 0.0042403629049658775, + 0.002840187167748809, + 0.003255313029512763, + 0.0035229516215622425, + 0.005948460195213556, + 0.004161816090345383, + 0.005137484986335039, + 0.0027507925406098366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.07899230718612671 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819830760359764 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.01819830760359764, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819830760359764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1232905313372612, + 0.08350455015897751, + 0.040217604488134384, + 0.12255485355854034, + 0.2251937985420227, + 0.06463322788476944, + 0.08568961173295975, + 0.1702745258808136, + 0.039072517305612564, + 0.04556886851787567 + ], + "credit_map": [ + 0.021405469626188278, + 0.007398017216473818, + 0.013377700932323933, + 0.007770885247737169, + 0.009036483243107796, + 0.01658986695110798, + 0.009083365090191364, + 0.03862874582409859, + 0.014794387854635715, + 0.003648902988061309, + 0.009166385047137737, + 0.005372367333620787, + 0.0049246326088905334, + 0.012006688863039017, + 0.007459348998963833, + 0.02613183669745922, + 0.006786506623029709, + 0.004158498719334602, + 0.005386922508478165, + 0.004291571211069822, + 0.002917259931564331, + 0.004834103863686323, + 0.00375520926900208, + 0.008087530732154846, + 0.01927832141518593, + 0.017346512526273727, + 0.012377974577248096, + 0.008618946187198162, + 0.015456822700798512, + 0.02215331420302391, + 0.01140289194881916, + 0.015920070931315422, + 0.058465033769607544, + 0.012211847119033337, + 0.013253355398774147, + 0.00983437430113554, + 0.013615213334560394, + 0.07859121263027191, + 0.009731481783092022, + 0.029491284862160683, + 0.013217803090810776, + 0.006357084959745407, + 0.008647065609693527, + 0.005761820822954178, + 0.0068706064485013485, + 0.007211536634713411, + 0.005428588483482599, + 0.01113872043788433, + 0.012783616781234741, + 0.0052856518886983395, + 0.019898414611816406, + 0.0038817087188363075, + 0.00472588837146759, + 0.007001311052590609, + 0.004187800921499729, + 0.027925215661525726, + 0.05431761592626572, + 0.006063386332243681, + 0.06224385276436806, + 0.009365968406200409, + 0.0072350045666098595, + 0.00986163504421711, + 0.007556951604783535, + 0.013630088418722153, + 0.004231299739331007, + 0.004812630359083414, + 0.00396704813465476, + 0.009155458770692348, + 0.0036558047868311405, + 0.003842692356556654, + 0.005854662973433733, + 0.003552922746166587, + 0.005607839208096266, + 0.005821503698825836, + 0.004095806274563074, + 0.0037341704592108727, + 0.009641718119382858, + 0.006962679326534271, + 0.006179468240588903, + 0.0035256813280284405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01819830760359764 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351071894168854 + ], + "opposite_logprob_tokens": [ + -3.651350975036621 + ], + "logprob_reward": -0.026351071894168854, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351071894168854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051356371492147446, + 0.036763936281204224, + 0.01513682771474123, + 0.5621731877326965, + 0.09466221928596497, + 0.026887385174632072, + 0.022297577932476997, + 0.14262333512306213, + 0.016000699251890182, + 0.032098401337862015 + ], + "credit_map": [ + 0.009490651078522205, + 0.0023817752953618765, + 0.006318284664303064, + 0.0033955536782741547, + 0.003221418708562851, + 0.006773322820663452, + 0.0039075626991689205, + 0.015867799520492554, + 0.0067656696774065495, + 0.0021150538232177496, + 0.0044192844070494175, + 0.002454169560223818, + 0.002370914677157998, + 0.006178117822855711, + 0.002856608247384429, + 0.009604117833077908, + 0.002933310577645898, + 0.0011594416573643684, + 0.002895057899877429, + 0.001558387652039528, + 0.0013564941473305225, + 0.0017051426693797112, + 0.0015049312496557832, + 0.0020240615122020245, + 0.10142870992422104, + 0.06965428590774536, + 0.04421631991863251, + 0.042719993740320206, + 0.0716068297624588, + 0.1284772753715515, + 0.042337965220212936, + 0.06173185259103775, + 0.024280373007059097, + 0.007502426393330097, + 0.005916468799114227, + 0.004205112345516682, + 0.006257571745663881, + 0.029950568452477455, + 0.0046874540857970715, + 0.011862244457006454, + 0.00522221066057682, + 0.002259923378005624, + 0.003891537431627512, + 0.0020124607253819704, + 0.0021056421101093292, + 0.00392976263538003, + 0.0020157163962721825, + 0.005450129508972168, + 0.0030496306717395782, + 0.0015057974960654974, + 0.0028210068121552467, + 0.003965151961892843, + 0.0018697476480156183, + 0.0034556814935058355, + 0.0027708625420928, + 0.002859698375687003, + 0.05354931205511093, + 0.005468434188514948, + 0.050346992909908295, + 0.007008700631558895, + 0.004185575060546398, + 0.006349985487759113, + 0.0036146710626780987, + 0.012099661864340305, + 0.0020978052634745836, + 0.0021083198953419924, + 0.0016371378442272544, + 0.0031577409245073795, + 0.0017234531696885824, + 0.0022751674987375736, + 0.001590543775819242, + 0.0014105309965088964, + 0.004352702293545008, + 0.0032493320759385824, + 0.002303631277754903, + 0.0024473785888403654, + 0.007879136130213737, + 0.0043420675210654736, + 0.005175461061298847, + 0.0023486940190196037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.026351071894168854 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796238988637924 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796238988637924, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796238988637924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05033025145530701, + 0.048830851912498474, + 0.0280915480107069, + 0.08192283660173416, + 0.13576604425907135, + 0.04941767826676369, + 0.05765921622514725, + 0.46376943588256836, + 0.037466924637556076, + 0.04674515128135681 + ], + "credit_map": [ + 0.009067825973033905, + 0.004488696809858084, + 0.007795616053044796, + 0.005546554923057556, + 0.004026568960398436, + 0.006915194448083639, + 0.0040743094868958, + 0.00841549038887024, + 0.008135640062391758, + 0.004510433413088322, + 0.008876274339854717, + 0.004111150745302439, + 0.003841671161353588, + 0.008713793940842152, + 0.003955502528697252, + 0.006686392240226269, + 0.005353668238967657, + 0.0021890278439968824, + 0.00486172828823328, + 0.0030074601527303457, + 0.002598855411633849, + 0.0039034849032759666, + 0.0024943214375525713, + 0.0036830033641308546, + 0.013518154621124268, + 0.009271778166294098, + 0.00815496128052473, + 0.006157808005809784, + 0.009686141274869442, + 0.01597452163696289, + 0.007142298389226198, + 0.012017179280519485, + 0.03835085406899452, + 0.00453551858663559, + 0.007448843214660883, + 0.004207082558423281, + 0.005642381496727467, + 0.05204285308718681, + 0.005626159254461527, + 0.017912374809384346, + 0.010985631495714188, + 0.004556283354759216, + 0.0064936853013932705, + 0.004294515587389469, + 0.004509475082159042, + 0.005735442973673344, + 0.00366782839410007, + 0.009174823760986328, + 0.007002365775406361, + 0.004560078959912062, + 0.01277413871139288, + 0.007482536137104034, + 0.005138893146067858, + 0.006836927030235529, + 0.005207845941185951, + 0.008656437508761883, + 0.15239502489566803, + 0.01718614064157009, + 0.18225926160812378, + 0.022691898047924042, + 0.014698044396936893, + 0.01773536205291748, + 0.01402873732149601, + 0.0427749939262867, + 0.005390456411987543, + 0.003895699279382825, + 0.005331718362867832, + 0.006912460085004568, + 0.0038322352338582277, + 0.005029402673244476, + 0.004042884800583124, + 0.0030320726800709963, + 0.0067109474912285805, + 0.006986591499298811, + 0.00524310115724802, + 0.0065604508854448795, + 0.0048359110951423645, + 0.007221807725727558, + 0.005250307731330395, + 0.003936039283871651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.029796238988637924 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sophisticated industrial product photography, a perfectly fixed macro close-up steadily captures the entire dynamic process. The background is an absolute, non-reflective black void, directing all visual focus to the central subject. Under bright, even commercial lighting, a round pressed powder sits motionless in the center of the frame. Its surface, a velvety matte, is embossed with a nebula-like pattern of saturated colors: vibrant magenta, deep cyan, and brilliant lemon yellow. A drop of crystal-clear, highly viscous liquid, like a perfect miniature glass sphere, slowly and precisely drips from the top of the frame. The moment the drop strikes the center of the powder, it instantly disrupts the powder's static equilibrium, triggering a miniature crown-shaped splash. The delicate, multicolored powders are then swept into the air, forming a dazzling and complex microscopic vortex. Driven by the liquid, these brilliant colors chase, swirl, stir, and merge with each other, like a silent color storm erupting on the millimeter scale. The whole process is full of scientific precision and stunning visual beauty, stably and clearly showing the dynamic interaction between materials.", + "seed": 9526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_9428_seed_9526_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": -0.2438894420862198, + "vlm_logprob": -0.2438894420862198, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.2438894420862198, + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9746930003166199 + ], + "opposite_logprob_tokens": [ + -0.4746930003166199 + ], + "logprob_reward": -0.9746930003166199, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9746930003166199, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11051253229379654, + 0.03571199253201485, + 0.041281312704086304, + 0.08001203835010529, + 0.08209248632192612, + 0.11757149547338486, + 0.3301161527633667, + 0.11961892247200012, + 0.04028937593102455, + 0.042793724685907364 + ], + "credit_map": [ + 0.01905127614736557, + 0.006597461644560099, + 0.017292877659201622, + 0.011597697623074055, + 0.006731292698532343, + 0.015507631003856659, + 0.010589035227894783, + 0.023145247250795364, + 0.006103778723627329, + 0.0035384269431233406, + 0.005910789594054222, + 0.003681381931528449, + 0.00467058876529336, + 0.004229356534779072, + 0.0033179239835590124, + 0.004259743262082338, + 0.007342912256717682, + 0.0044016363099217415, + 0.0057184030301868916, + 0.003913624212145805, + 0.006431418936699629, + 0.004880343098193407, + 0.0036222755443304777, + 0.004970697220414877, + 0.02618596702814102, + 0.005417047068476677, + 0.02569265477359295, + 0.005539047531783581, + 0.0038664962630718946, + 0.005419319495558739, + 0.0031667372677475214, + 0.004724767990410328, + 0.015794342383742332, + 0.0064340694807469845, + 0.015592277981340885, + 0.008391888812184334, + 0.007188580464571714, + 0.01119045540690422, + 0.008679972030222416, + 0.008820896036922932, + 0.022717634215950966, + 0.009655935689806938, + 0.0212540403008461, + 0.012761665508151054, + 0.008770821616053581, + 0.014658574014902115, + 0.011415677145123482, + 0.016337141394615173, + 0.062472160905599594, + 0.029854534193873405, + 0.035712406039237976, + 0.025862036272883415, + 0.029492385685443878, + 0.05612468719482422, + 0.03652166947722435, + 0.054076261818408966, + 0.035729315131902695, + 0.01198784913867712, + 0.016454145312309265, + 0.010042264126241207, + 0.008487910032272339, + 0.009272078983485699, + 0.010373160243034363, + 0.017272183671593666, + 0.009508868679404259, + 0.0043535842560231686, + 0.004481795243918896, + 0.004007704555988312, + 0.002953803399577737, + 0.004454998765140772, + 0.004379946272820234, + 0.006148672662675381, + 0.007420482113957405, + 0.0029288821388036013, + 0.006880941800773144, + 0.003980668261647224, + 0.003708804724738002, + 0.004921106155961752, + 0.005851884372532368, + 0.007100952789187431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.9746930003166199 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032293945550919 + ], + "opposite_logprob_tokens": [ + -2.350322961807251 + ], + "logprob_reward": -0.10032293945550919, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032293945550919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1563149392604828, + 0.04289581626653671, + 0.04669297859072685, + 0.17734697461128235, + 0.0553731694817543, + 0.10200405865907669, + 0.15232032537460327, + 0.16296997666358948, + 0.029978780075907707, + 0.0741029754281044 + ], + "credit_map": [ + 0.02834359183907509, + 0.014942309819161892, + 0.02056914195418358, + 0.02046668902039528, + 0.014584716409444809, + 0.01738225854933262, + 0.020694009959697723, + 0.019332217052578926, + 0.007558662444353104, + 0.004011209588497877, + 0.005369368940591812, + 0.004514405038207769, + 0.006039788015186787, + 0.00453098863363266, + 0.0052824970334768295, + 0.005588893312960863, + 0.009673605673015118, + 0.005657241214066744, + 0.007080488372594118, + 0.004815035033971071, + 0.006148934364318848, + 0.0040799411945044994, + 0.004979010205715895, + 0.0042587160132825375, + 0.06322292238473892, + 0.008312972262501717, + 0.058970071375370026, + 0.011489275842905045, + 0.0057299998588860035, + 0.011143028736114502, + 0.00798677746206522, + 0.010491910390555859, + 0.011613325215876102, + 0.004501729737967253, + 0.0095703499391675, + 0.005370431113988161, + 0.0031786488834768534, + 0.0067989155650138855, + 0.004495123866945505, + 0.00984464306384325, + 0.014573480002582073, + 0.006877109408378601, + 0.01687094196677208, + 0.009849552996456623, + 0.008981924504041672, + 0.014659657143056393, + 0.011069069616496563, + 0.01912231743335724, + 0.02984234131872654, + 0.013936713337898254, + 0.017329728230834007, + 0.012162785977125168, + 0.013283513486385345, + 0.025164082646369934, + 0.0177448783069849, + 0.022856269031763077, + 0.02980477176606655, + 0.02195274457335472, + 0.01749403402209282, + 0.019778108224272728, + 0.01483371201902628, + 0.015308370813727379, + 0.021344423294067383, + 0.022453803569078445, + 0.005281892139464617, + 0.003062913892790675, + 0.004378675017505884, + 0.004047751426696777, + 0.002564751310274005, + 0.0026176937390118837, + 0.004146633204072714, + 0.0038784693460911512, + 0.012716242112219334, + 0.006151661276817322, + 0.01292604673653841, + 0.008664287626743317, + 0.007716686464846134, + 0.00809276383370161, + 0.007672719657421112, + 0.010162560269236565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.10032293945550919 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370916470885277 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.03370916470885277, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370916470885277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.241595059633255, + 0.05156669765710831, + 0.08304513245820999, + 0.06970089673995972, + 0.0981738269329071, + 0.07878294587135315, + 0.07164035737514496, + 0.1464015245437622, + 0.06555511057376862, + 0.09353850781917572 + ], + "credit_map": [ + 0.042707882821559906, + 0.029740065336227417, + 0.03156992420554161, + 0.023795414716005325, + 0.029728621244430542, + 0.02392556518316269, + 0.024430086836218834, + 0.035697486251592636, + 0.008261769078671932, + 0.004416807554662228, + 0.007017242256551981, + 0.004100901540368795, + 0.005772868636995554, + 0.006881425157189369, + 0.005473204422742128, + 0.009642479009926319, + 0.016034938395023346, + 0.007701187860220671, + 0.011803822591900826, + 0.011432047002017498, + 0.009950496256351471, + 0.008047808893024921, + 0.009520183317363262, + 0.00855464581400156, + 0.008861472830176353, + 0.0076655796729028225, + 0.011133069172501564, + 0.009634656831622124, + 0.008494225330650806, + 0.007686176802963018, + 0.006580041721463203, + 0.00964567344635725, + 0.023194128647446632, + 0.008667847141623497, + 0.016841381788253784, + 0.010992223396897316, + 0.005352309439331293, + 0.01170619111508131, + 0.009857210330665112, + 0.011562532745301723, + 0.01158895157277584, + 0.005940043367445469, + 0.014581947587430477, + 0.007395580876618624, + 0.007123254705220461, + 0.00977917481213808, + 0.006655034609138966, + 0.01571895368397236, + 0.011015268974006176, + 0.006529088132083416, + 0.008253827691078186, + 0.009077589027583599, + 0.005341021344065666, + 0.010908778756856918, + 0.01138676144182682, + 0.009128022007644176, + 0.030246661975979805, + 0.017002474516630173, + 0.016013270244002342, + 0.01927839033305645, + 0.012153424322605133, + 0.01457629632204771, + 0.016201600432395935, + 0.020929399877786636, + 0.010781978257000446, + 0.007301101461052895, + 0.011051991023123264, + 0.00584851810708642, + 0.005589242093265057, + 0.0074310097843408585, + 0.00738098006695509, + 0.01017028745263815, + 0.01935448683798313, + 0.0055854301899671555, + 0.013139193877577782, + 0.007179178297519684, + 0.0058753760531544685, + 0.00930747203528881, + 0.009368881583213806, + 0.023728493601083755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.03370916470885277 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048671357333660126 + ], + "opposite_logprob_tokens": [ + -3.048671245574951 + ], + "logprob_reward": -0.048671357333660126, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048671357333660126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17990276217460632, + 0.04672820493578911, + 0.04801994562149048, + 0.24694637954235077, + 0.11793390661478043, + 0.06554780155420303, + 0.07617049664258957, + 0.04439380764961243, + 0.025888808071613312, + 0.14846782386302948 + ], + "credit_map": [ + 0.03606433793902397, + 0.0156333576887846, + 0.028064748272299767, + 0.024172788485884666, + 0.01893870159983635, + 0.015843799337744713, + 0.02166002057492733, + 0.01952502317726612, + 0.007689429447054863, + 0.0033422140404582024, + 0.005919028539210558, + 0.006636134348809719, + 0.004363718442618847, + 0.0066592018119990826, + 0.0044598267413675785, + 0.0076586585491895676, + 0.008715247735381126, + 0.006482391152530909, + 0.006165431346744299, + 0.003499103244394064, + 0.010262658819556236, + 0.005140076857060194, + 0.00309614441357553, + 0.004658897873014212, + 0.0923205018043518, + 0.010449642315506935, + 0.10021482408046722, + 0.008481314405798912, + 0.0057936906814575195, + 0.015601438470184803, + 0.005472118034958839, + 0.008612866513431072, + 0.03070046566426754, + 0.009777304716408253, + 0.023395780473947525, + 0.00872354581952095, + 0.005272120703011751, + 0.01708614081144333, + 0.008462127298116684, + 0.014516427181661129, + 0.011387854814529419, + 0.005336117465049028, + 0.014617087319493294, + 0.00597138237208128, + 0.004277344793081284, + 0.008340144529938698, + 0.005551837384700775, + 0.010066033340990543, + 0.011505519039928913, + 0.006429390516132116, + 0.008746675215661526, + 0.0062171039171516895, + 0.007297743111848831, + 0.014617177657783031, + 0.008504172787070274, + 0.01285272091627121, + 0.007289132568985224, + 0.005568698979914188, + 0.004067494999617338, + 0.00436689518392086, + 0.004012551624327898, + 0.005233134608715773, + 0.005566126201301813, + 0.008289780467748642, + 0.004491996020078659, + 0.0027596713043749332, + 0.004244175273925066, + 0.002591178985312581, + 0.0018765667919069529, + 0.0024557935539633036, + 0.0029584842268377542, + 0.004510945174843073, + 0.024431223049759865, + 0.009574608877301216, + 0.029261885210871696, + 0.01323745772242546, + 0.011498757638037205, + 0.020151738077402115, + 0.015897098928689957, + 0.02441507577896118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.048671357333660126 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062050748616456985 + ], + "opposite_logprob_tokens": [ + -2.8120508193969727 + ], + "logprob_reward": -0.062050748616456985, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062050748616456985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522444188594818, + 0.045274894684553146, + 0.05521835759282112, + 0.15913845598697662, + 0.0777389407157898, + 0.04089128226041794, + 0.15716344118118286, + 0.0733998566865921, + 0.06693591922521591, + 0.1719944030046463 + ], + "credit_map": [ + 0.028601156547665596, + 0.011026368476450443, + 0.020291335880756378, + 0.016587816178798676, + 0.011469969525933266, + 0.021269317716360092, + 0.015486182644963264, + 0.027512267231941223, + 0.007489243987947702, + 0.0041817850433290005, + 0.007027046289294958, + 0.005191977135837078, + 0.006020626984536648, + 0.004859483800828457, + 0.004724259488284588, + 0.005780471488833427, + 0.009874175302684307, + 0.006039950996637344, + 0.006171150133013725, + 0.007108405232429504, + 0.006200585979968309, + 0.006285855080932379, + 0.005342776887118816, + 0.008195457980036736, + 0.056305088102817535, + 0.009764756076037884, + 0.05736415088176727, + 0.008280142210423946, + 0.005910961888730526, + 0.008416734635829926, + 0.004655221477150917, + 0.008441407233476639, + 0.016446372494101524, + 0.006210621912032366, + 0.009424499236047268, + 0.010288018733263016, + 0.007172850426286459, + 0.009774596430361271, + 0.00791352242231369, + 0.010508458130061626, + 0.006570490077137947, + 0.00341890100389719, + 0.006721400190144777, + 0.0038959896191954613, + 0.004028412513434887, + 0.005690803751349449, + 0.0035616597160696983, + 0.007003627251833677, + 0.027740733698010445, + 0.01459087710827589, + 0.014514748938381672, + 0.011039161123335361, + 0.015232190489768982, + 0.028728507459163666, + 0.01957705244421959, + 0.02574017085134983, + 0.010955546982586384, + 0.008600888773798943, + 0.00609795656055212, + 0.009155607782304287, + 0.007029416970908642, + 0.0072369505651295185, + 0.01034583617001772, + 0.013977648690342903, + 0.010783064179122448, + 0.009731722995638847, + 0.010888184420764446, + 0.006203261204063892, + 0.005023361649364233, + 0.006434028968214989, + 0.00574870640411973, + 0.012123589403927326, + 0.027063356712460518, + 0.011589525267481804, + 0.035436250269412994, + 0.01822921261191368, + 0.014401191845536232, + 0.022772440686821938, + 0.01848280057311058, + 0.024019628763198853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.062050748616456985 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sophisticated industrial product photography, a perfectly fixed macro close-up steadily captures the entire dynamic process. The background is an absolute, non-reflective black void, directing all visual focus to the central subject. Under bright, even commercial lighting, a round pressed powder sits motionless in the center of the frame. Its surface, a velvety matte, is embossed with a nebula-like pattern of saturated colors: vibrant magenta, deep cyan, and brilliant lemon yellow. A drop of crystal-clear, highly viscous liquid, like a perfect miniature glass sphere, slowly and precisely drips from the top of the frame. The moment the drop strikes the center of the powder, it instantly disrupts the powder's static equilibrium, triggering a miniature crown-shaped splash. The delicate, multicolored powders are then swept into the air, forming a dazzling and complex microscopic vortex. Driven by the liquid, these brilliant colors chase, swirl, stir, and merge with each other, like a silent color storm erupting on the millimeter scale. The whole process is full of scientific precision and stunning visual beauty, stably and clearly showing the dynamic interaction between materials.", + "seed": 9527, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_9428_seed_9527_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": -0.15018625631928445, + "vlm_logprob": -0.15018625631928445, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.15018625631928445, + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5766302347183228 + ], + "opposite_logprob_tokens": [ + -0.8266302347183228 + ], + "logprob_reward": -0.5766302347183228, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5766302347183228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0803392305970192, + 0.017252158373594284, + 0.025943558663129807, + 0.036894239485263824, + 0.2141437530517578, + 0.3318744897842407, + 0.20339879393577576, + 0.02968357689678669, + 0.034231215715408325, + 0.02623903378844261 + ], + "credit_map": [ + 0.01883823610842228, + 0.007081115618348122, + 0.010772431269288063, + 0.007385424803942442, + 0.005333939101547003, + 0.012376513332128525, + 0.009072564542293549, + 0.009479005821049213, + 0.0027083850000053644, + 0.001384016708470881, + 0.00320827541872859, + 0.001751916715875268, + 0.0011748450342565775, + 0.0020320864859968424, + 0.00225144624710083, + 0.0027411864139139652, + 0.008271219208836555, + 0.0020783194340765476, + 0.002689109416678548, + 0.002291671000421047, + 0.001806023414246738, + 0.003552981885150075, + 0.002183384494856, + 0.00307084945961833, + 0.010975253768265247, + 0.002756212605163455, + 0.003228735411539674, + 0.002705329330638051, + 0.0024764982517808676, + 0.00649925647303462, + 0.002797196852043271, + 0.0054557593539357185, + 0.05143937095999718, + 0.01626339927315712, + 0.01779315061867237, + 0.02099677547812462, + 0.01390176359564066, + 0.038170404732227325, + 0.017217637971043587, + 0.03836124762892723, + 0.1366899013519287, + 0.020045002922415733, + 0.013959954492747784, + 0.01352947298437357, + 0.01933000423014164, + 0.06163300573825836, + 0.013203554786741734, + 0.05348360165953636, + 0.07537689059972763, + 0.010093947872519493, + 0.006882404442876577, + 0.010216702707111835, + 0.006631619296967983, + 0.07586482167243958, + 0.008645601570606232, + 0.00968680065125227, + 0.004164367448538542, + 0.006079443264752626, + 0.0033399339299649, + 0.003536578267812729, + 0.002855063183233142, + 0.003594973124563694, + 0.0029928849544376135, + 0.0031203334219753742, + 0.0044651636853814125, + 0.0035318725276738405, + 0.004240061622112989, + 0.00595899997279048, + 0.002946918597444892, + 0.00393603416159749, + 0.004990153945982456, + 0.004162008874118328, + 0.0039016958326101303, + 0.0047794608399271965, + 0.0017286718357354403, + 0.002250856254249811, + 0.0036070607602596283, + 0.00413175392895937, + 0.0027031071949750185, + 0.00313642923720181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.5766302347183228 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002826035022736 + ], + "opposite_logprob_tokens": [ + -2.695028305053711 + ], + "logprob_reward": -0.07002826035022736, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002826035022736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09130533039569855, + 0.05115840584039688, + 0.04713243618607521, + 0.027270909398794174, + 0.06705543398857117, + 0.2583794593811035, + 0.20957690477371216, + 0.06327594071626663, + 0.07721153646707535, + 0.10763365030288696 + ], + "credit_map": [ + 0.014313998632133007, + 0.009221798740327358, + 0.010786422528326511, + 0.009559059515595436, + 0.008179759606719017, + 0.009725004434585571, + 0.01698269695043564, + 0.012536592781543732, + 0.008850154466927052, + 0.006917131133377552, + 0.00993905495852232, + 0.005892827175557613, + 0.0043959347531199455, + 0.0039061731658875942, + 0.0058149611577391624, + 0.005442166700959206, + 0.016932623460888863, + 0.003713863203302026, + 0.0035827725660055876, + 0.004309727344661951, + 0.0026748618111014366, + 0.008011287078261375, + 0.002604485722258687, + 0.005302822683006525, + 0.005011390428990126, + 0.0016261773416772485, + 0.00261942227371037, + 0.0021196226589381695, + 0.0021042542066425085, + 0.007585411891341209, + 0.0027133121620863676, + 0.0034913206472992897, + 0.011717185378074646, + 0.006957965437322855, + 0.006934065371751785, + 0.007988400757312775, + 0.005073952954262495, + 0.011657369323074818, + 0.0063498737290501595, + 0.010376626625657082, + 0.10179974883794785, + 0.021298911422491074, + 0.011156965978443623, + 0.011550623923540115, + 0.017522180452942848, + 0.049503907561302185, + 0.009719480760395527, + 0.03582765534520149, + 0.06542463600635529, + 0.01725592464208603, + 0.013023697771131992, + 0.01262590941041708, + 0.011153814382851124, + 0.06582608819007874, + 0.014058287255465984, + 0.010208552703261375, + 0.00879704486578703, + 0.010979666374623775, + 0.006147872190922499, + 0.0060443938709795475, + 0.00536196306347847, + 0.012802976183593273, + 0.0048697046004235744, + 0.008272320963442326, + 0.013424055650830269, + 0.011137673631310463, + 0.00781184621155262, + 0.008238434791564941, + 0.007968844845890999, + 0.009974019601941109, + 0.010182640515267849, + 0.008474032394587994, + 0.02158469147980213, + 0.008307678624987602, + 0.007888248190283775, + 0.007841210812330246, + 0.0068373833782970905, + 0.03209846839308739, + 0.009182383306324482, + 0.013893600553274155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.07002826035022736 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181954987347126 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.0181954987347126, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181954987347126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10297506302595139, + 0.06401050090789795, + 0.02907979115843773, + 0.021999584510922432, + 0.06326575577259064, + 0.08754090964794159, + 0.32753610610961914, + 0.043214745819568634, + 0.05143466591835022, + 0.20894289016723633 + ], + "credit_map": [ + 0.014754547737538815, + 0.007632417604327202, + 0.01428097765892744, + 0.014492320828139782, + 0.008592495694756508, + 0.008957886137068272, + 0.020644349977374077, + 0.013620071113109589, + 0.01161434967070818, + 0.005590794142335653, + 0.007291126996278763, + 0.008063072338700294, + 0.0046594892628490925, + 0.009085242636501789, + 0.009432961232960224, + 0.008273462764918804, + 0.008902499452233315, + 0.002347395522519946, + 0.0029500843957066536, + 0.0030573515687137842, + 0.0013551040319725871, + 0.0050152940675616264, + 0.0019638906233012676, + 0.003488171612843871, + 0.003831361187621951, + 0.0013644818682223558, + 0.002449748106300831, + 0.002907464047893882, + 0.0017192419618368149, + 0.005174674093723297, + 0.002598122926428914, + 0.001954491715878248, + 0.019949186593294144, + 0.0046083685010671616, + 0.004329704213887453, + 0.005444130394607782, + 0.0035796796437352896, + 0.009536813013255596, + 0.004396896809339523, + 0.01142097171396017, + 0.022871606051921844, + 0.01399548351764679, + 0.005908382125198841, + 0.006786962505429983, + 0.009599181823432446, + 0.01385294459760189, + 0.005694562569260597, + 0.008831790648400784, + 0.12609481811523438, + 0.017057182267308235, + 0.01018291711807251, + 0.013180907815694809, + 0.011309153400361538, + 0.12291575223207474, + 0.012011256068944931, + 0.014784112572669983, + 0.005478335544466972, + 0.008092832751572132, + 0.003941361326724291, + 0.0048550087958574295, + 0.0038906405679881573, + 0.008844828233122826, + 0.003833241993561387, + 0.004278496373444796, + 0.010022363625466824, + 0.005009918473660946, + 0.006139763630926609, + 0.007017410825937986, + 0.006844354793429375, + 0.004832009319216013, + 0.00667581707239151, + 0.004893029574304819, + 0.039514631032943726, + 0.020791277289390564, + 0.015386125072836876, + 0.01573757268488407, + 0.015374928712844849, + 0.055452071130275726, + 0.019673019647598267, + 0.027013247832655907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0181954987347126 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08116026967763901, + 0.04396160691976547, + 0.05657774582505226, + 0.022945651784539223, + 0.02584194950759411, + 0.013144169934093952, + 0.5898627042770386, + 0.056800685822963715, + 0.04769715294241905, + 0.06200804561376572 + ], + "credit_map": [ + 0.01822914183139801, + 0.007637535221874714, + 0.010440940037369728, + 0.007842155173420906, + 0.005611777305603027, + 0.0114420922473073, + 0.010360202752053738, + 0.009596421383321285, + 0.008534979075193405, + 0.004025407135486603, + 0.006416280288249254, + 0.006109326612204313, + 0.0030284621752798557, + 0.005158531479537487, + 0.005695831961929798, + 0.004992785397917032, + 0.00978227611631155, + 0.004922195803374052, + 0.005652047228068113, + 0.009541270323097706, + 0.005900444928556681, + 0.0062123797833919525, + 0.008718305267393589, + 0.005848823115229607, + 0.005187097936868668, + 0.00195510545745492, + 0.0020009081345051527, + 0.0028441885951906443, + 0.0016546061960980296, + 0.0030418646056205034, + 0.0021131837274879217, + 0.004148694220930338, + 0.003201402025297284, + 0.003905977588146925, + 0.0029559831600636244, + 0.0034957951866090298, + 0.0025646386202424765, + 0.0033377825748175383, + 0.0032306744251400232, + 0.0031496926676481962, + 0.0019425192149356008, + 0.002288192044943571, + 0.0013976751361042261, + 0.001927521894685924, + 0.0018620886839926243, + 0.0016233822098001838, + 0.001042848452925682, + 0.0010599405504763126, + 0.21994183957576752, + 0.02823607623577118, + 0.015823327004909515, + 0.021208645775914192, + 0.01730925776064396, + 0.23526717722415924, + 0.021002905443310738, + 0.03107348643243313, + 0.007071726955473423, + 0.009589551016688347, + 0.006685060914605856, + 0.007584350183606148, + 0.0056287674233317375, + 0.0078062270767986774, + 0.005900113843381405, + 0.006534879561513662, + 0.008594668470323086, + 0.0048802257515490055, + 0.00546587398275733, + 0.00504506379365921, + 0.005346771329641342, + 0.005835726857185364, + 0.00689059030264616, + 0.005638227332383394, + 0.010883395560085773, + 0.01000560075044632, + 0.004096030257642269, + 0.0063557871617376804, + 0.0069005535915493965, + 0.012124944478273392, + 0.005344417877495289, + 0.0062973108142614365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000903785228729 + ], + "opposite_logprob_tokens": [ + -2.6950089931488037 + ], + "logprob_reward": -0.07000903785228729, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000903785228729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11799657344818115, + 0.025506310164928436, + 0.03748992085456848, + 0.013037669472396374, + 0.06666237115859985, + 0.06456693261861801, + 0.4675854742527008, + 0.07605038583278656, + 0.07081806659698486, + 0.0602862723171711 + ], + "credit_map": [ + 0.024003705009818077, + 0.009295228868722916, + 0.021576447412371635, + 0.013524236157536507, + 0.010727209970355034, + 0.014124799519777298, + 0.013121981173753738, + 0.011622967198491096, + 0.00445083063095808, + 0.0028692029882222414, + 0.0032653934322297573, + 0.0028650222811847925, + 0.002224920317530632, + 0.002825750969350338, + 0.0034336792305111885, + 0.00357151054777205, + 0.006584278307855129, + 0.0033103295136243105, + 0.004251813981682062, + 0.006655441131442785, + 0.0031841180752962828, + 0.00489916792139411, + 0.0047168307937681675, + 0.0038879411295056343, + 0.002413759008049965, + 0.0009764551068656147, + 0.0013159794034436345, + 0.0009988404344767332, + 0.001093546743504703, + 0.002751536900177598, + 0.0013128151185810566, + 0.002174736699089408, + 0.024683430790901184, + 0.005981702357530594, + 0.0037481999024748802, + 0.004459448158740997, + 0.003342684591189027, + 0.010010749101638794, + 0.003102255519479513, + 0.011333896778523922, + 0.020719585940241814, + 0.008188381791114807, + 0.003869430162012577, + 0.0046989512629806995, + 0.0057819318026304245, + 0.011856473051011562, + 0.0030456585809588432, + 0.006406518165022135, + 0.17018121480941772, + 0.024943092837929726, + 0.01714358851313591, + 0.02532365545630455, + 0.017337780445814133, + 0.1642453372478485, + 0.025264952331781387, + 0.023145848885178566, + 0.011985446326434612, + 0.013553861528635025, + 0.007471729069948196, + 0.008594867773354053, + 0.00612659240141511, + 0.011578164994716644, + 0.008718305267393589, + 0.008021415211260319, + 0.011037840507924557, + 0.010208338499069214, + 0.005296835210174322, + 0.015654053539037704, + 0.006885179318487644, + 0.008283348754048347, + 0.006641576066613197, + 0.0068108937703073025, + 0.011766886338591576, + 0.005500433500856161, + 0.004217455629259348, + 0.0036331312730908394, + 0.0037935846485197544, + 0.019185027107596397, + 0.0041145626455545425, + 0.00807519443333149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.07000903785228729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of a sophisticated industrial product photography, a perfectly fixed macro close-up steadily captures the entire dynamic process. The background is an absolute, non-reflective black void, directing all visual focus to the central subject. Under bright, even commercial lighting, a round pressed powder sits motionless in the center of the frame. Its surface, a velvety matte, is embossed with a nebula-like pattern of saturated colors: vibrant magenta, deep cyan, and brilliant lemon yellow. A drop of crystal-clear, highly viscous liquid, like a perfect miniature glass sphere, slowly and precisely drips from the top of the frame. The moment the drop strikes the center of the powder, it instantly disrupts the powder's static equilibrium, triggering a miniature crown-shaped splash. The delicate, multicolored powders are then swept into the air, forming a dazzling and complex microscopic vortex. Driven by the liquid, these brilliant colors chase, swirl, stir, and merge with each other, like a silent color storm erupting on the millimeter scale. The whole process is full of scientific precision and stunning visual beauty, stably and clearly showing the dynamic interaction between materials.", + "seed": 9528, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000167/group_9/index_9428_seed_9528_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": -0.20893817264586687, + "vlm_logprob": -0.20893817264586687, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.20893817264586687, + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8989199995994568 + ], + "opposite_logprob_tokens": [ + -0.5239199995994568 + ], + "logprob_reward": -0.8989199995994568, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8989199995994568, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20643207430839539, + 0.04631628841161728, + 0.13983528316020966, + 0.039907507598400116, + 0.04219457879662514, + 0.25391092896461487, + 0.07343055307865143, + 0.0719488188624382, + 0.09311491250991821, + 0.03290910646319389 + ], + "credit_map": [ + 0.06496056914329529, + 0.01108354888856411, + 0.012714911252260208, + 0.015611277893185616, + 0.016500603407621384, + 0.06398029625415802, + 0.010834957472980022, + 0.010745909996330738, + 0.006779862102121115, + 0.00327794230543077, + 0.009459417313337326, + 0.004664410836994648, + 0.0045478143729269505, + 0.006992731709033251, + 0.0036467667669057846, + 0.006947341840714216, + 0.039921682327985764, + 0.012021658942103386, + 0.008863362483680248, + 0.009641223587095737, + 0.02262135036289692, + 0.03165069594979286, + 0.005956004839390516, + 0.00915930513292551, + 0.006349447648972273, + 0.0035857639741152525, + 0.005777763202786446, + 0.0037946226075291634, + 0.005384764168411493, + 0.0055577936582267284, + 0.0027786255814135075, + 0.006678726524114609, + 0.006198883056640625, + 0.0050204722210764885, + 0.004327951930463314, + 0.004068812821060419, + 0.005219647660851479, + 0.005448669660836458, + 0.005284932907670736, + 0.0066252113319933414, + 0.008629340678453445, + 0.019000070169568062, + 0.011106045916676521, + 0.006016172468662262, + 0.011553147807717323, + 0.0985424667596817, + 0.007393239997327328, + 0.09167042374610901, + 0.007592839654535055, + 0.009612026624381542, + 0.007812615483999252, + 0.009969675913453102, + 0.006481605116277933, + 0.01175607368350029, + 0.011865495704114437, + 0.008340215310454369, + 0.0074331737123429775, + 0.014290064573287964, + 0.007165323942899704, + 0.010446435771882534, + 0.007787494454532862, + 0.009474537335336208, + 0.006694153416901827, + 0.008657634258270264, + 0.011147965677082539, + 0.010601820424199104, + 0.012761182151734829, + 0.01060525979846716, + 0.010580260306596756, + 0.011470461264252663, + 0.010993887670338154, + 0.014954076148569584, + 0.003396885469555855, + 0.00413226755335927, + 0.0042928121984004974, + 0.0037810674402862787, + 0.004564488772302866, + 0.0029850516002625227, + 0.004510337952524424, + 0.005246195010840893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.8989199995994568 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905178308486938 + ], + "opposite_logprob_tokens": [ + -2.4640517234802246 + ], + "logprob_reward": -0.08905178308486938, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905178308486938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15577413141727448, + 0.11512957513332367, + 0.10296063870191574, + 0.04052146524190903, + 0.04178767278790474, + 0.01779637672007084, + 0.027918921783566475, + 0.021619506180286407, + 0.04195677861571312, + 0.4345349371433258 + ], + "credit_map": [ + 0.032966382801532745, + 0.00868330616503954, + 0.01095388364046812, + 0.012171926908195019, + 0.009855854324996471, + 0.04041460156440735, + 0.010855639353394508, + 0.029872525483369827, + 0.020420972257852554, + 0.0053884368389844894, + 0.007477563805878162, + 0.007116657216101885, + 0.0049334801733493805, + 0.021108705550432205, + 0.007411250378936529, + 0.04127250611782074, + 0.026771195232868195, + 0.008953792974352837, + 0.007505824323743582, + 0.007739170454442501, + 0.015936238691210747, + 0.022848811000585556, + 0.005877371411770582, + 0.00732823321595788, + 0.0058219158090651035, + 0.0049918582662940025, + 0.005012352950870991, + 0.004010334145277739, + 0.008093440905213356, + 0.004243785515427589, + 0.003499438287690282, + 0.004848336335271597, + 0.006955069489777088, + 0.005803295411169529, + 0.004494457505643368, + 0.004470250569283962, + 0.00629688985645771, + 0.005048439837992191, + 0.004689099732786417, + 0.004030169919133186, + 0.0015742360847070813, + 0.003361383220180869, + 0.0021007165778428316, + 0.0017418002244085073, + 0.002955886535346508, + 0.0023392345756292343, + 0.001677931286394596, + 0.0020451874006539583, + 0.002730375388637185, + 0.003493963275104761, + 0.002849097829312086, + 0.003201838815584779, + 0.0044244760647416115, + 0.004550924990326166, + 0.0038609246257692575, + 0.002807322423905134, + 0.002122344449162483, + 0.0030573648400604725, + 0.0026023704558610916, + 0.003643171628937125, + 0.0018382384441792965, + 0.002560489345341921, + 0.003494416130706668, + 0.0023011111188679934, + 0.0056236214004457, + 0.0064471703954041, + 0.004809877835214138, + 0.005214380566030741, + 0.004816337488591671, + 0.004888205323368311, + 0.004324386361986399, + 0.005832802038639784, + 0.024217035621404648, + 0.02923431620001793, + 0.023533688858151436, + 0.021675104275345802, + 0.10445017367601395, + 0.025453953072428703, + 0.14594145119190216, + 0.06002918630838394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.08905178308486938 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357457041740417 + ], + "opposite_logprob_tokens": [ + -3.651357412338257 + ], + "logprob_reward": -0.026357457041740417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357457041740417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17227803170681, + 0.10757390409708023, + 0.21130645275115967, + 0.057232122868299484, + 0.08327832072973251, + 0.049349330365657806, + 0.05547505244612694, + 0.04601604491472244, + 0.12046989798545837, + 0.09702079743146896 + ], + "credit_map": [ + 0.031681280583143234, + 0.010769560001790524, + 0.015473982319235802, + 0.015712862834334373, + 0.02140737697482109, + 0.03884097561240196, + 0.01127505861222744, + 0.02711694873869419, + 0.01601535826921463, + 0.007548911031335592, + 0.011161979287862778, + 0.016175972297787666, + 0.010693876072764397, + 0.01611495390534401, + 0.010392065159976482, + 0.019470790401101112, + 0.049723464995622635, + 0.019245091825723648, + 0.014194176532328129, + 0.016451487317681313, + 0.034626416862010956, + 0.04533732309937477, + 0.011107739992439747, + 0.02062077447772026, + 0.008016831241548061, + 0.00532549899071455, + 0.010003899224102497, + 0.007890268228948116, + 0.008085966110229492, + 0.006125327665358782, + 0.004773795139044523, + 0.007010540924966335, + 0.015731872990727425, + 0.008062651380896568, + 0.00617927685379982, + 0.0073175085708498955, + 0.009390203282237053, + 0.010638035833835602, + 0.007387649733573198, + 0.01857113093137741, + 0.0036359219811856747, + 0.009501528926193714, + 0.004205598961561918, + 0.0062031131237745285, + 0.007886999286711216, + 0.006750582251697779, + 0.0047377608716487885, + 0.006427828222513199, + 0.006131927017122507, + 0.010922432877123356, + 0.004155340604484081, + 0.006599050015211105, + 0.005974522791802883, + 0.008135959506034851, + 0.006947730202227831, + 0.006608095020055771, + 0.003939531743526459, + 0.005395805928856134, + 0.0064340317621827126, + 0.006969199515879154, + 0.0036355070769786835, + 0.007260204292833805, + 0.007334334775805473, + 0.0050474368035793304, + 0.01560359075665474, + 0.022978896275162697, + 0.01319834403693676, + 0.014734688214957714, + 0.012435219250619411, + 0.016462896019220352, + 0.0103948088362813, + 0.014661461114883423, + 0.005312873516231775, + 0.007686328608542681, + 0.006147188134491444, + 0.006973587907850742, + 0.02206190675497055, + 0.006246958393603563, + 0.029501982033252716, + 0.0130899827927351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.026357457041740417 + }, + { + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14176678657531738, + 0.03780845180153847, + 0.08405903726816177, + 0.07544170320034027, + 0.057539086788892746, + 0.14046025276184082, + 0.019535822793841362, + 0.016205817461013794, + 0.02743673510849476, + 0.39974626898765564 + ], + "credit_map": [ + 0.02431699074804783, + 0.00683941226452589, + 0.01183276902884245, + 0.009909787215292454, + 0.009133807383477688, + 0.031647637486457825, + 0.007701897993683815, + 0.04038449376821518, + 0.006049433723092079, + 0.0021052046213299036, + 0.00342111405916512, + 0.002870422787964344, + 0.0030688780825585127, + 0.005761928856372833, + 0.0022547224070876837, + 0.012276747263967991, + 0.011412161402404308, + 0.010337607003748417, + 0.007788588292896748, + 0.010940687730908394, + 0.01906672865152359, + 0.008595758117735386, + 0.006709988694638014, + 0.00920751504600048, + 0.012767781503498554, + 0.010115467011928558, + 0.006136198528110981, + 0.005887824576348066, + 0.016917532309889793, + 0.007779981009662151, + 0.004722787998616695, + 0.011114130727946758, + 0.00967223197221756, + 0.009670644998550415, + 0.0060374317690730095, + 0.006641114596277475, + 0.008441376499831676, + 0.005493890028446913, + 0.004676982294768095, + 0.006905411370098591, + 0.005110718309879303, + 0.011164424009621143, + 0.00636107474565506, + 0.0038517580833286047, + 0.00847627967596054, + 0.05165805295109749, + 0.0047834813594818115, + 0.04905445873737335, + 0.0016427453374490142, + 0.0026833422016352415, + 0.0017851508455350995, + 0.0014513981295749545, + 0.003625771962106228, + 0.004867966752499342, + 0.001305883633904159, + 0.0021735636983066797, + 0.0017599635757505894, + 0.0037229189183562994, + 0.0018990564858540893, + 0.002132772235199809, + 0.0013917058240622282, + 0.00185316929128021, + 0.0016077018808573484, + 0.001838531345129013, + 0.00269881309941411, + 0.003970936872065067, + 0.002960361074656248, + 0.004877317231148481, + 0.0028247092850506306, + 0.002678776392713189, + 0.0032488375436514616, + 0.00417698360979557, + 0.022393520921468735, + 0.0261420626193285, + 0.017528750002384186, + 0.017846157774329185, + 0.09562330693006516, + 0.0223673265427351, + 0.1408703327178955, + 0.05697481334209442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "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 background an absolute, non-reflective black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a round pressed powder with a velvety matte surface visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the powder surface feature an embossed pattern of vibrant magenta, deep cyan, and brilliant lemon yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a drop of crystal-clear, highly viscous liquid drip from the top and strike the powder to create a crown-shaped splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the powders form a microscopic vortex that swirls and merges into a complex, multicolored pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11082254350185394, + 0.10233809798955917, + 0.09192010760307312, + 0.03437945246696472, + 0.0439126156270504, + 0.06235583871603012, + 0.037977252155542374, + 0.042007818818092346, + 0.3333180844783783, + 0.1409681886434555 + ], + "credit_map": [ + 0.031860098242759705, + 0.005962898954749107, + 0.007060487288981676, + 0.007896780967712402, + 0.0069970483891665936, + 0.03185677528381348, + 0.005133384373039007, + 0.014055070467293262, + 0.019345546141266823, + 0.0046985396184027195, + 0.006070369854569435, + 0.003969273995608091, + 0.004505912307649851, + 0.020244169980287552, + 0.005773332901299, + 0.037730954587459564, + 0.021680327132344246, + 0.008046790957450867, + 0.005975003819912672, + 0.006712683010846376, + 0.013075287453830242, + 0.018227888271212578, + 0.0067487191408872604, + 0.011453405022621155, + 0.00566150201484561, + 0.004967996850609779, + 0.0027146462816745043, + 0.0032235796097666025, + 0.007670002523809671, + 0.0035868906415998936, + 0.00220518303103745, + 0.004349649418145418, + 0.005724393762648106, + 0.007413775660097599, + 0.004521599505096674, + 0.0036343189422041178, + 0.007306783925741911, + 0.00460274051874876, + 0.003417345928028226, + 0.007291655987501144, + 0.005248656962066889, + 0.010709441266953945, + 0.005211205687373877, + 0.005529412534087896, + 0.009210853837430477, + 0.011095285415649414, + 0.0062811849638819695, + 0.009069794788956642, + 0.004536898341029882, + 0.0070879533886909485, + 0.005545464344322681, + 0.003978236112743616, + 0.004011016339063644, + 0.005253571085631847, + 0.00318347942084074, + 0.004380632191896439, + 0.004720679018646479, + 0.0049545480869710445, + 0.006120538804680109, + 0.004991248715668917, + 0.004770295694470406, + 0.0049146381206810474, + 0.005023238714784384, + 0.0065126302652060986, + 0.04093029350042343, + 0.0464947447180748, + 0.03746752813458443, + 0.05159864202141762, + 0.03535991534590721, + 0.04799310863018036, + 0.03660392761230469, + 0.03686990961432457, + 0.008921625092625618, + 0.008527467958629131, + 0.011012797243893147, + 0.010603870265185833, + 0.030756700783967972, + 0.00879489816725254, + 0.04378340020775795, + 0.01856742426753044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020579153671860695 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file